Thursday, September 5, 2019

* Registration - held at DoubleTree by Hilton Hotel Cluj – City Plaza, 9-13 Sindicatelor Street, Cluj-Napoca, Romania.
** Welcome Party - held at DoubleTree by Hilton Hotel Cluj – City Plaza, Cluj-Napoca, Romania, Marco Polo Restaurant

Friday, September 6, 2019

8:00: Registration
DoubleTree by Hilton Hotel Cluj – City Plaza, 9-13 Sindicatelor Street, Cluj-Napoca, Romania.

*Banquet - held at DoubleTree by Hilton Hotel Cluj – City Plaza - Marco Polo Restaurant.

Saturday, September 7, 2019

8:00: Registration
DoubleTree by Hilton Hotel Cluj – City Plaza, 9-13 Sindicatelor Street, Cluj-Napoca, Romania.

Detailed Technical Program

Workshops

Deep Learning in Perception and Control (Thursday, September 5, 14:00 - 15:50)
Location: Beijing Room, 5^th floor

Workshop Details: here.

Bosch Student Workshop In Automated Driving (Thursday, September 5, 16:10 - 18:00)
Location: Beijing Room, 5^th floor

Workshop Details: here.

Keynote Lectures

Plenary Presentation 1 (Friday, September 6, 09:00 - 09:50)
Location: Ballroom, 1^st floor

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

Technical University of Cluj-Napoca, Romania
Technical University of Cluj-Napoca, Romania

Virtual integration and large coverage testing of automotive functions
Mugur Tatar	QTronic GmbH, Berlin, Germany

Plenary Presentation 2 (Friday, September 6, 10:10 - 11:00)
Location: Ballroom, 1^st floor

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

Technical University of Cluj-Napoca, Romania
Technical University of Cluj-Napoca, Romania

Deep Learning for Autonomous Driving
Luis M. Bergasa	University of Alcalá, Spain

Plenary Presentation 3 (Friday, September 6, 11:20 - 12:10)
Location: Ballroom, 1^st floor

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

Technical University of Cluj-Napoca, Romania
Technical University of Cluj-Napoca, Romania

Self-Driving Vehicles in the City: The Vulnerable Road User Challenge
Dariu M. Gavrila	TU Delft, The Netherlands

Conference Sessions

Special Session: Automated Driving 1 (Friday, September 6, 13:30 - 14:50)
Location: Ballroom , 1^st Floor

Chair: Pasi Pyykönen
Co-Chair: Florin Oniga

VTT Technical Research Centre of Finland Ltd., Finland
Technical University of Cluj-Napoca, Romania

13:30 - 13:50
Obstacle Detection Using a Voxel Octree Representation View Paper
Sorana Capalnean Florin Oniga Radu Danescu	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: This paper proposes a novel clustering algorithm based on an octree representation, suitable for stereovision-based riving assistance applications. The voxel octree representation takes into account the stereovision perspective and uncertainty model, thus dealing better with the 3D raw measurements. Given the internal partitioning of the octree, the indexing of children nodes can be used directly to determine 3D adjacent leaves. Therefore, the proposed clustering scheme exploits the octree structure to establish the 3D connectivity of voxels and detect 3D obstacle primitives. Computational complexity of clustering is lowered this way. These primitives are then merged based on adjacency criteria in order to obtain the 3D obstacles. KITTI’s adtaset was used to evaluate the accuracy of the approach. The proposed algorithm is applicable to k-ary trees for any dimensionality.

13:50 - 14:10
Multi-Object tracking of 3D cuboids using aggregated feature channels View Paper
Mircea Paul Muresan Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The unknown correspondences of measurements and targets, referred to as data association, is one of the main challenges of multi-target tracking. Each new measurement received could be the continuation of some previously detected target, the first detection of a new target or a false alarm. Tracking 3D cuboids, is particularly difficult due to the high amount of data, which can include erroneous or noisy information coming from sensors, that can lead to false measurements, detections from an unknown number of objects which may not be consistent over frames or varying object properties like dimension and orientation. In the self-driving car context, the target tracking module holds an important role due to the fact that the ego vehicle has to make predictions regarding the position and velocity of the surrounding objects in the next time epoch, plan for actions and make the correct decisions. To tackle the above mentioned problems and other issues coming from the self-driving car processing pipeline we propose three original contributions: 1) designing a novel affinity measurement function to associate measurements and targets using multiple types of features coming from LIDAR and camera, 2) a context aware descriptor for 3D objects that improves the data association process, 3) a framework that includes a module for tracking dimensions and orientation of objects. The implemented solution runs in real time and experiments that were performed on real world urban scenarios prove that the presented method is effective and robust even in a highly dynamic environment

14:10 - 14:30
Estimating pedestrian intentions from trajectory data View Paper
Elena Sikudova Kristina Malinovska Radoslav Skoviera Julia Skovierova Miroslav Uller Vaclav Hlavac	CIIRC, CVUT Prague, Czechia CIIRC, CVUT Prague, Czechia CIIRC, CVUT Prague, Czechia CIIRC, CVUT Prague, Czechia CIIRC, CVUT Prague, Czechia CIIRC, CVUT Prague, Czechia
Abstract: Intention estimation of traffic participants is the most crucial part of autonomous driving systems. Especially important in urban situations is the intention estimation of pedestrians. In this paper, several machine learning methods are trained and evaluated in the task of estimation of the intention of a pedestrian to cross a zebra crossing. Their results are compared to a Bayesian network – an approach commonly used in autonomous driving. The data used for the estimation contain only position and heading of the pedestrians. The best performing method achieved F 2 score of 92.37.

14:30 - 14:50
LiDAR Performance Review in Arctic Conditions View Paper
Maria Jokela Matti Kutila Kimmo Kauvo Pasi Pyykönen	VTT Technical Research Centre of Finland Ltd., Finland VTT Technical Research Centre of Finland Ltd., Finland VTT Technical Research Centre of Finland Ltd., Finland VTT Technical Research Centre of Finland Ltd., Finland
Abstract: This article aims to outline the key results of testing and encountered challenges of various LiDARs, radar and stereo camera in arctic weather conditions. The test session was conducted in two different urban areas in Finland in the middle of January 2019. The arctic conditions turned out to be challenging for the sensors dedicated more to areas where temperature stays relatively warm. The aim of this one-week test session was to assess performance deterioration when powdered snow, salted road, snowy ground and sun light influence reliability of the future automated driving functions. This study focuses mainly on the issues with hardware that are basic building blocks for the situation awareness software modules. Furthermore, the countermeasures such as protecting sensors and mounting positions have been proposed. The test results indicate that some sensors significantly lose performance when temperature drops to less than -10 degrees centigrade. The problem is not merely mechanical freezing of the spinning LiDAR components but properties of laser illumination may change due to temperature variation, too. Since LiDAR is an optical device, they also suffer when there is turbulent snow in front of the sensor. The turbulence looks like a noise and partially blocks the laser echoes from surrounding environment. The performance can with some sensors drop more than 50 percent. This seriously diminishes the sensing range and furthermore, makes pattern recognition unreliable. The two other sensor types which were taken into account are stereo vision and radar. They have a role in automated driving to compensate performance degradation of LiDARs due to arctic conditions.

Data Driven Robotic Intelligent Systems Workshop (Friday, September 6, 13:30 - 14:50)
Location: Venezia Room, 5^th Floor

Chair: Adina Magda Florea
Co-Chair: Rodica Potolea

University Politehnica of Bucharest, Romania
Technical University of Cluj-Napoca, Romania

13:30 - 13:50
End-to-end models for self-driving cars on UPB campus roads View Paper
Andrei Mihalea Robert Samoilescu Andrei Cristian Nica Mihai Trăscău Alexandru Sorici Adina Magda Florea	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania
Abstract: End-to-end learning using deep neural networks for autonomous driving problems has been approached using various methods and techniques including supervised learning, unsupervised learning and reinforcement learning. The goal is to provide a model able to produce steering, break and acceleration commands for a car by using as input 2D images captured by on-board cameras. We describe a process to design and train such models given a specific geographical area with its own particularities in terms of road network, traffic and driving conditions. We propose and data augmentation techniques which improve the performance of our models. Based on a collected dataset in the UPB campus we compared the performance of several of our proposed models trained directly on this data employing several augmentation techniques. Evaluation is performed using our proposed simulation pipeline which apply transformations to the frames in test videos such that they reflect decisions made by the model in earlier steps. This brings the system evaluation closer to real conditions. Results indicate the proposed models offer a good performance for the collected dataset.

13:50 - 14:10
The Impact of Data Challenges on Intent Detection and Slot Filling for the Home Assistant Scenario View Paper
Anda Stoica Tibor Kadar Camelia Lemnaru Rodica Potolea Mihaela Dinsoreanu	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Natural Language Understanding (NLU) is currently a very high-interest domain to both academia and the commercial environment, due in the largest part to the recent increased popularity of conversational systems. In this paper we focus on the home assistant application context and identify a set of language and data-related challenges that can occur in such a scenario, such as: distribution shift, missing information and class imbalance. We systematically generate datasets in the Romanian language that model these data complexities and further investigate how well two of the most prominent tools - Wit.ai and Rasa NLU - solve the tasks of intent detection and slot filling, given the considered data complexities. We perform a thorough analysis of the errors produced by both tools, and provide the most probable justification for their occurrence. We found that both tools focus extensively on the verb for identifying intents, and that antonyms, class-imbalance and certain small variations in formulation greatly impact intent and slot identification. This opens new research directions to directly address these shortcomings.

14:10 - 14:30
Intelligent control of an aerodynamical system View Paper
Alex Danku Andrei Kovari Liviu C. Miclea Eva-Henrietta Dulf	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The paper presents the designed prototype for a highly nonlinear, multi-input-multi-output aerodynamic system. The laboratory scale equipment is created to simulate the operations of unmanned aerial vehicles. It was conceived to be cheap and easy to use, in order to be multiplied for laboratory works. It is also described the first tested control strategy, based on dynamic nonlinear model inversion using artificial neural networks. The experimental results prove the efficiency of the equipment, being able to test different real operation behaviors.

14:30 - 14:50
Keyword Spotting using Dynamic Time Warping and Convolutional Recurrent Networks View Paper
Erika-Timea Albert Camelia Lemnaru Mihaela Dinsoreanu Rodica Potolea	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: This paper proposes a method for keyword spotting, which first converts utterances to grayscale images via a modified Dynamic Time Warping (DTW) algorithm, and then splits the images into frames which are fed in sequence to a Convolutional Recurrent Deep Neural Network (CRDNN). DTW is employed because of its capability to accurately capture similarities between time sequences, while the neural network exploits the textural features of the DTW matrix for classification. We explore three alternative formulations of the DTW algorithm for extracting the similarity matrices, as well as three different conversion methods from the similarity matrix to a grayscale image. As opposed to previous works, we employ a recurrent network to consider sequential information encoded in image segments. We perform several evaluations on the TIMIT corpus and find that the system reaches a detection performance of 95%.

Special Session: HiPerGRID (Friday, September 6, 13:30 - 14:50)
Location: Beijing Room, 5^th Floor

Chair: Florin Pop

Co-Chair: Dorian Gorgan

University Politehnica of Bucharest, Romania
National Institute for Research and Development
in Informatics (ICI), Bucharest, Romania
Technical University of Cluj-Napoca, Romania

13:30 - 13:50
Distributed processing models for large unstructured datasets: satellite imagery usecase View Paper
Ion Dorinel Filip Cătălin Negru Florin Pop Adrian Stoica Florin Șerban	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania National Institute for Research and Development in Informatics (ICI), Bucharest, Romania Terrasigna, Romania Terrasigna, Romania
Abstract: For most of the programs, the most significant amount of time is spent on running a CPU-intensive component, while retrieving and reading the input or formatting the output only takes an insignificant amount of time, especially if both the input and output are structured data. Thus, most of the computer science literature on distributed systems discuss the optimization of algorithms and frequently treat the I/O and input preparation parts of the program execution as a much less important one for overall efficiency. That well-informed decision of ignoring the I/O part when discussing runtime optimization might change to a fault when we consider algorithms running over massive datasets that should also be retrieved at a user’s request. In this paper, we propose and analyze an user-centric processing platform for running algorithms over satellite imagery products, including a review of the important challenges, solutions, and opportunities on extracting valuable results from GIS products. The 4th section of the paper includes two examples of real-life usage of remote sensing for the observation of natural habitats. In the 5th section we include three practical experiments assessing different aspects of the proposed solution.

13:50 - 14:10
Predicting Service Level Agreement Violations in Cloud using Machine Learning techniques View Paper
George Iordache Florin Pop	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania National Institute for Research and Development in Informatics (ICI), Bucharest, Romania
Abstract: When discussing about Service Level Agreement contract design it is very important to deal with violations in the Cloud. This means that if the scheduling process doesn’t take in consideration some workloads (it cannot schedule them) this will generate issues in the interaction between Cloud Service Customer(s) and the Cloud Service Provider. In our article we deal with estimating the actual workload that is sent to be scheduled on the Cloud Service Infrastructure (CSI). The estimation is based on two different machine learning techniques: Hidden Markov Models and Neural Networks. We chose two different learning techniques to show that workload prediction can be done in various ways. The results show with a certain confidence that can be considered good enough we can predict the actual load on the Cloud Service Infrastructure.

14:10 - 14:30
Building soil classification maps using HorusApp and Sentinel-2 Products View Paper
Victor Bacu Teodor Stefanut Dorian Gorgan	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: An efficient agricultural management is based on a good, reach and accurate information on the environment and especially on the soil. The need for up-to-date and high-resolution soil information and the direct access to this information in a flexible and simple manner is imperative for pedology and agriculture specialists. This paper presents the HORUSApp application supporting the integration of multispectral data coming from satellite images (in particular from Sentinel-2 satellite) into the soil analysis and classification process.

14:30 - 14:50
Usability evaluation of a domain-specific language for defining aggregated processing tasks View Paper
Constantin Ioan Nandra Dorian Gorgan	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The effective processing of Big Data sets often requires some programming knowledge from a prospective user’s part. This could prove costly to achieve, in terms of user training time and effort, depending on the level of previous experience. The premise, when dealing with large data sets, is that it should be as easy as possible for a user to prototype and test processing algorithms, in order to deal with them in an effective manner. For this reason, we have developed a domain-specific language meant to allow users to define data processing tasks as aggregates, consisting of atomic operations. Its goal is to do away with some of the complexities of traditional programming languages, by simplifying the representation model and providing a more intuitive process description tool for its users. This paper aims to evaluate the efficiency and effectiveness with which a novice user could employ our domain-specific language to define processing tasks, and then compare the results to those obtained while using the Python programming language. The experiments will be focused on task duration, description correctness and code interpretation, highlighting possible advantages and disadvantages observed during the usage of the two languages.

Special Session: Automated Driving 2 (Friday, September 6, 15:05 - 16:25)
Location: Ballroom , 1^st Floor

Chair: Matti Kutila
Co-Chair: Mihai Negru

VTT Technical Research Centre of Finland Ltd., Finland
Technical University of Cluj-Napoca, Romania

15:05 - 15:25
Sensor System Power Adaptation for Automated Vehicles View Paper
Matti Kutila Ari Virtanen Mikko Tarkiainen Pertti Peussa	VTT Technical Research Centre of Finland Ltd., Finland VTT Technical Research Center of Finland Ltd., Finland VTT Technical Research Center of Finland Ltd., Finland VTT Technical Research Center of Finland Ltd., Finland
Abstract: There are two mega-trends in the automotive industry for converting the way of traveling in the future. One is the transition from combustion engines to electric vehicles and the second promotes an increased level of automation in future cars. However, automation requires exceptionally high quality environment perception data and various sensor devices, each taking their own share from the vehicle energy storage, which reduces the driving range of electric vehicles. This article proposes an improved ability to save the energy of a vehicle through advanced sensor systems with an adaptive power management of the environment perception system. At low speeds the sensor system is set to a minimal power scheme and at high speed or in adverse weather conditions a high power scheme is used to increase the available sensing range. In addition, it also makes it possible to use V2X communication to restrict sensor power or prohibit the use of certain sensors in radiation sensitive environments.

15:25 - 15:45
Toward a hybrid vehicle communication platform based on VLC and DSRC technologies View Paper
Eduard Zadobrischi Sebastian Andrei Avătămăniței Alin Mihai Căilean Mihai Dimian Mihai Negru	Stefan cel Mare University of Suceava, Romania Stefan cel Mare University of Suceava, Romania Stefan cel Mare University of Suceava, Romania Stefan cel Mare University of Suceava, Romania Technical University of Cluj-Napoca, Romania
Abstract: Wireless communications technologies have the potential to significantly improve the safety and the efficiency of the transportation system. Nevertheless, supporting the requirements imposed by the usage in vehicle safety applications is rather challenging. In such a context, this paper provides an overview of some of the most promising technologies that could be used in communication-based vehicle safety applications. The paper provides an overview of the DSRC technology as a favorite technology for such applications and of the VLC technology as a very promising emergent alternative. Based on this analysis, a hybrid architecture using DSRC and VLC is proposed. This architecture envisions that neighboring vehicles are interconnected based on VLC, whereas non-line of sight vehicles are connected through the DSRC network. Thus, an application for the connection of these network solutions is proposed and evaluated.

15:45 - 16:05
Rapid Light Field Depth Estimation with Semi-Global Matching View Paper
Yuriy Anisimov Oliver Wasenmüller Didier Stricker	German Research Centre for Artificial Intelligence, Germany German Research Centre for Artificial Intelligence, Germany German Research Centre for Artificial Intelligence, Germany
Abstract: Running time of the light field depth estimation algorithms is typically high. This assessment based on the computational complexity of existing methods and the large amounts of data involved. The aim of our work is to develop a simple and fast algorithm for accurate depth computation. In this context, we propose an approach, which involves Semi-Global Matching for the processing of light field images. It forms on comparison of pixels' correspondences with different metrics in the substantially bounded light field space. We show that our method is suitable for the quick production of a proper result in a variety of light field configurations.

16:05 - 16:25
Automatic Extrinsic Calibration of LIDAR and Monocular Camera Images View Paper
Albert-Szabolcs Vaida Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Mobile robots are being equipped with an ever increasing amount of sensors, and by taking advantage of each modality's forte, through sensor fusion we are able to produce superior results for higher level functions such as pedestrian detection. An accurate calibration between the sensors, however, is essential for obtaining a correct fusion. One of the more common setups is that of an optical camera paired with a lidar system, complementing each other's data. While the camera offers rich color information through a dense image, the lidar produces high accuracy, though sparse depth measurements. In this paper we will present two methods for estimating the 6 degrees of freedom parameters necessary to extrinsically calibrate the two sensors, without the use of special calibration targets: the first one attempts to align edges detected in both the lidar point cloud and the corresponding color image, while the second approach minimizes the depth difference between the measured lidar data and a depth map generated from the corresponding monocular image. By defining correlations between the two modality's information, we are able to construct an optimization problem on the 6 calibration parameters. Our experiments show good results when matched against similar solutions proposed in the literature, especially in the case of the calibration algorithm based on depth estimation.

Knowledge and Ontology Engineering (Friday, September 6, 15:05 - 16:25)
Location: Venezia Room, 5^th Floor

Chair: Christophe Nicolle
Co-Chair: Emil Stefan Chifu

Université de Bourgogne Franche-Comté, France
Technical University of Cluj-Napoca, Romania

15:05 - 15:25
Enterprise Knowledge Modeling, UML vs ontology: Formal Evaluation View Paper
Meriem Mejhed Mkhinini Ouassila Labbani Christophe Nicolle	Université de Bourgogne Franche-Comté, France Université de Bourgogne Franche-Comté, France Université de Bourgogne Franche-Comté, France
Abstract: Everyday activities in enterprises rely heavily on the experts' know-how. Due to experts departure, the loss of expertise and knowledge is a reoccurring problem in these enterprises. Recently, in order to capture experts knowledge into intelligent systems, formal knowledge representation methods, such as ontologies, are being studied and have caught up with non-formal or semi-formal representation, such as UML. The similarities and differences between UML class diagram and computational ontology have for long raised questions about the possibility of synthesizing them in a common representation (usually an ontology). Indeed, the problem of migrating knowledge encoded in UML into an ontology is an active research domain. Here, we present our approach, which is based on semantic matching between existing ontologies and a UML class diagram, to support UML driven ontology refactoring and engineering.

15:25 - 15:45
Identifying Concepts and Relations in a Transition-Based AMR Parser View Paper
Viorel Ieremias Roxana Pop Florin Macicasan	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: AMR is a graph-based semantic representation for natural language. It relies on concepts and their relations to transcend words and distill the meaning of English sentences. In this work, we propose a solution for identifying both the concepts and their associated relations as a post-processing step of a transition-based parsing system. We incorporate these contributions within an existing system. Furthermore, we enhance the LSTM transition learning component by optimizing the input features and extending the predicted output.

15:45 - 16:05
Sentiment Analysis of Events in Social Media View Paper
Ciprian-Octavian Truică Elena Apostol Alexandru Petrescu	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania
Abstract: The growing popularity of Online Social Networks has open new research directions and perspectives for content analysis, i.e., Network Analysis and Natural Language Processing. From the perspective of information spread, the Network Analysis community propose Event Detection. This approach focuses on the network features, without an in-depth analysis of the textual content, summarization being a preferred method. Natural Language Processing analyses only the textual content, not integrating the graph-based structure of the network. To address these limitations, we propose a method that bridges the two directions and integrates content-awareness into network-awareness. Our method uses event detection to extract topics of interest and then applies sentiment analysis on each event. The obtained results have high accuracy, proving that our method determines with high precision the overall sentiment of the detected events.

16:05 - 16:25
An Unsupervised Neural Model for Aspect Based Opinion Mining View Paper
Emil Stefan Chifu Viorica Rozina Chifu	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The paper approaches aspect based opinion mining, with uses an unsupervised neural network as the opinion classifier. The neural network is an extension of the Growing Hierarchical Self-organizing Maps (GHSOM). In the aspect based sentiment analysis, we exploit the fact that the binary relations of syntactic dependency, extracted from product reviews, very often express relations between an aspect of the reviewed product and an opinion towards that aspect. We use the Growing Hierarchical Self-organizing Maps to classify pairs built according to the dependency relations, more exactly to classify pairs of the form (aspect, opinion bearing word). With this classification, we discover whether the various text mentions of the aspects of the target entity (such as the aspects of a product) are opinionated with positive or negative sentiment in the text of a review. We classify these pairs against a domain specific taxonomy of aspects, which also includes (positive/ negative) opinions associated with the aspects. Since it is based on classification against an ontology, our approach is semantic oriented. We tested our system on a collection of reviews about photo cameras.

Distributed Systems 1 (Friday, September 6, 15:05 - 16:25)
Location: Beijing Room, 5^th Floor

Chair: Adrian Coleșa
Co-Chair: Tudor Cioara

Technical University of Cluj-Napoca, Romania
Technical University of Cluj-Napoca, Romania

15:05 - 15:25
Securely Exposing Machine Learning Models to Web Clients using Intel SGX View Paper
Dávid Ács Adrian Coleșa	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Machine learning (ML) methods are applied frequently to predict outcomes or features, that would otherwise require tedious manual work. ML models are usually deployed on Web servers, where end user can query them providing the input data. Server side deployment's shortcoming is that users' data must be sent to a server on each query, increasing network usage and leading to privacy/legal issues. In this paper we present a system which aims to ease the deployment of ML models on the client side of Web applications, while protecting the he Intellectual Property (IP) of the model owner. Protection of the ML model is realized with Intel SGX which assures that a loaded model cannot be inspected by the end-user.

15:25 - 15:45
Host Oriented Factor Normalizing Authentication Resource View Paper
Aurel-Dragos Hofnar Marius Joldos	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Whenever one accesses a computer system there are three essential security issues involved: identification, authentication and authorization. The identification process enables recognition of an entity, which may be either a human, a machine, or another asset – e.g. software program. Two complementary mechanisms are used for determining who can access those systems: authentication and authorization. To address the authentication process, various solutions have been proposed in the literature, from a simple password to newer technologies based on biometrics or RFID (Radio Frequency IDentification). This paper presents a novel scalable multi-factor authentication method, applicable to computer systems with no need of any hardware/software changes.

15:45 - 16:05
Detection of Instability Sources in Smart Grids Using Machine Learning Techniques View Paper
Dorin Moldovan Ioan Salomie	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The prediction of smart grid stability represents a challenging research problem because this information can be very useful for the identification of the participants that lead to instabilities in the system and consequently it is very useful to determine configurations in which the system is stable even if some participants present abnormalities. The main contributions of this research article are: (1) the presentation of a machine learning methodology for predicting the smart grid stability which is based on features extraction using the tsfresh package from Python, (2) the selection of features using three methods namely, Binary Particle Swarm Optimization Features Selection (BPSOFS), Binary Kangaroo Mob Optimization Features Selection (BKMOFS) and Multivariate Adaptive Regression Splines (MARS), (3) the prediction of the system's stability using four classifiers namely, Logistic Regression (LR), Random Forest (RF), Gradient Boosted Trees (GBT) and Multilayer Perceptron Classifier (MPC) and (4) the detection of instability sources using a method based on machine learning and statistics. The best prediction results are obtained when MPC is applied (93.8%) and when the features are selected using BPSOFS.

16:05 - 16:25
A Stacking Multi-Learning Ensemble Model for Predicting Near Real Time Energy Consumption Demand of Residential Buildings View Paper
Andreea Valeria Vesa Nicoleta Ghitescu Claudia Pop Marcel Antal Tudor Cioara Ionut Anghel Ioan Salomie	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The aim of this paper is to present a novel energy consumption forecasting solution for predicting energy demand at the level of residential buildings based on their historical consumption profile for a seamless integration with the session-based Energy Markets developed within Smart Energy Grids that integrate renewable energy sources. To overcome the drawbacks and lack of accuracy of existing prediction models, a stacked multi-learning ensemble model is proposed combining Gradient Boosting Regression, Multi-Layer Neural Networks and Long Short Term Memory Networks followed by a Linear Regressor for forecasting residential energy demands both at individual and aggregated levels. The proposed ensemble predictor is evaluated using the open-access UK-DALE dataset containing historical energy traces for 5 households spreading over several years, obtaining a best MAPE of 1.59%, a RMSE of 6.19kWh and a MAE of 5.60kWh on the aggregated dataset, proving the high accuracy and stability of the proposed solution as well as the feasibility of using ensemble models for residential building energy demand forecast for integration with session-based energy marketplaces.

Special Session: Automated Driving 3 (Friday, September 6, 16:40 - 18:00)
Location: Ballroom , 1^st Floor

Chair: Ana L. C. Bazzan
Co-Chair: Radu Danescu

Universidade Federal do Rio Grande do Sul, Brazil
Technical University of Cluj-Napoca, Romania

16:40 - 17:00
Dynamic 3D Environment Perception Using Monocular Vision and Semantic Segmentation View Paper
Radu Danescu Razvan Itu Diana Borza	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: This paper presents a complete system for traffic environment perception based on a single color image source. The acquired color images are segmented into road and obstacle areas using a U-Net style Convolutional Neural Network (CNN). The segmented image is mapped into a bird’s-eye view using the automatically calibrated camera parameters. Obstacle scans are extracted from the bird’s-eye view image, highlighting the contact points between the obstacles and the road. The scans are post-processed to increase the connectivity between obstacle parts. The processed scans are used to generate the measurement likelihood values for all observable cells of a dynamic occupancy grid, taking into consideration the expected measurement errors with respect to the distance from the camera. A particle-based occupancy grid is used to track the environment at cell level, and then the occupied cells are grouped into individual objects. The system is able to estimate stable cuboids with measured width, length, orientation and speed for the moving individual objects such as vehicles and pedestrians, and also to identify generic occupied areas for the continuous structures such as fences or barriers.

17:00 - 17:20
Personalized Gear Shifting Architecture for Next Generation Automatic Transmission Systems View Paper
Ayşegül Sarı Ahmet Taha Bilgiç Görkem Şafak Duygu Erates Emre Kaplan	AVL Research and Engineering Istanbul, Turkey, Turkey AVL Research and Engineering Istanbul, Turkey, Turkey AVL Research and Engineering Istanbul, Turkey, Turkey AVL Research and Engineering Istanbul, Turkey, Turkey AVL Research and Engineering Istanbul, Turkey, Turkey
Abstract: Personalization is one of the trending topics of nowadays. Artiﬁcial Intelligence based technologies enable us to personalize systems to reﬂect user desire and driving proﬁle. In automotive domain, we see intelligent software takes place in many aspects of the vehicle including transmission systems. Today, most of the vehicles are produced with an automatic transmission system which works as programmed according to the development expertise but does not incorporate behavior feedbacks. This paper proposes a novel contribution to automatic transmission systems by incorporating driver feedback to achieve personalization. This way, next generation automatic transmission systems can learn from user behavior taking their inputs into account and reﬂect under certain conditions. The system learns driver’s demands on the road via supervised learning and predicts driver’s desired gear according to the road conditions, user manipulations, and all relevant information gathered from the vehicle at run time. Learning desires of the driver can ﬁt into the automatic transmission’s decision-making process without violating safety standards and the operational durability as well as leaving very small footprint in terms of memory, space and computation respecting to the limited capability of the environment that the method resides. The proposed method was tested in a realistic testing environment and the results are promising so that it can be deployed in a vehicle to extend automatic transmissions’ capabilities with personalization. In fact, personalized shifting leads to better customer experience and retention.

17:20 - 17:40
Object Detection in Monocular Infrared Images Using Classification – Regresion Deep Learning Architectures View Paper
Raluca Brehar Flaviu Vancea Tiberiu Marita Cristian Vancea Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical Unversity of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The rapid development of deep learning architectures that have a good performance on object detection in visual monocular images has triggered and interest towards the application of these architectures on other image modalities such as stereovision or infrared images. We propose a framework for multi-class object detection in monocular infrared images that integrates and compares different classification-regression deep learning architectures on a novel benchmark infrared dataset developed by FLIR. The work described is evaluated using standard object detection metrics and an average precision of 82% for pedestrians, 86% for cars and 66% for bicycles is achieved while running at 40fps.

17:40 - 18:00
Using Information from Heterogeneous Sources and Machine Learning in Intelligent Transportation Systems View Paper
Ana L. C. Bazzan Jorge Chamby-Diaz Rhuam Sena Estevam Leonardo de A. Schmidt Marcia Pasin Jorge L. A. Samatelo Matheus Vieira Lessa Ribeiro	Universidade Federal do Rio Grande do Sul, Brazil Universidade Federal do Rio Grande do Sul, Brazil Universidade Federal do Rio Grande do Sul, Brazil Universidade Federal de Santa Maria, Brazil Universidade Federal de Santa Maria, Brazil Universidade Federal do Espirito Santo, Brazil Universidade Federal do Espirito Santo, Brazil
Abstract: Current estimates show that in Brazil traffic causes losses of around ten billion BRL per year. Although Intelligent Transportation Systems (ITS) techniques can contribute to reduce this figure, many of the proposed ITS-related techniques have not been yet developed. This is especially the case of techniques that use the potential of the Internet. Thus, this paper proposes methods for retrieving and using data from heterogeneous sources, currently available on the Internet, in order to provide information for both authorities and logistics services, as well as for the citizen, with Machine Learning (ML) support. It is important to stress that such Internet data may not only be traffic-related but also refer to other sources such as text (social networks such as Twitter); meteorological bulletins; sports and cultural events; images of traffic flow; videos (webcams); inter-vehicular communication and other sources connected with mobility Internet. Underlying this effort, we propose a framework in which ML techniques play a key role in several tasks, from image and text processing to prediction. Specifically, our applications deal with named entities extraction in tweets, pothole detection, and classification of traffic state from images.

Natural Language Processing (Friday, September 6, 16:40 - 18:00)
Location: Venezia Room, 5^th Floor

Chair: Ciprian Dobre

Co-Chair: Radu Razvan Slavescu

National Institute for Research and Development
in Informatics (ICI), Bucharest, Romania
Technical University of Cluj-Napoca, Romania

16:40 - 17:00
Romanian Part of Speech Tagging using LSTM Networks View Paper
Beata Lorincz Maria Nutu Adriana Stan	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: In this paper we present LSTM based neural network architectures for determining the part of speech (POS) tags for Romanian words. LSTM networks combined with fully-connected output layers are used for predicting the root POS, and sequence-to-sequence models composed of LSTM encoders and decoders are evaluated for predicting the extended MSD and CTAG tags. The highest accuracy achieved for the root POS is 99.18% and for the extended tags is 98.25%. This method proves to be efficient for the proposed task and has the advantage of being language independent, as no expert linguistic knowledge is used in the input features.

17:00 - 17:20
A New Challenge in the Data Processing of Non-Standard Texts Containing Accents / Diacritics: A Case Study View Paper
Veronica Gavrila Lidia Bajenaru Ciprian Dobre	National Institute for Research and Development in Informatics (ICI), Bucharest, Romania National Institute for Research and Development in Informatics (ICI), Bucharest, Romania National Institute for Research and Development in Informatics (ICI), Bucharest, Romania
Abstract: The INTELLIT project develops a virtual online library of the Romanian literature. The sources of data made available and provided by the Romanian Academy, such as: General Dictionary of Romanian Literature, Timeline of the Romanian Literary Life and the canonical works of Romanian writers are digitized and indexed using smart text analytics. One of the challenges with this process is dealing with diacritics and textual accents. Here, we present an in-depth analysis of possible solutions and describe our implementation for detecting various Unicode text processing. We present the solution identified as an accessible way to remove specific Unicode text code points in order to greatly improve our search and filtering capabilities while still preserving the original source (at the database level).

17:20 - 17:40
Deep Learning for Automatic Diacritics Restoration in Romanian View Paper
Maria Nuțu Beata Lorincz Adriana Stan	Babes-Bolyai University, Romania Babes-Bolyai University, Romania Technical University of Cluj-Napoca, Romania
Abstract: In this paper we address the issue of automatic diacritics restoration (ADR) for Romanian using deep learning strategies. We compare 6 separate architectures with various mixtures of recurrent and convolutional layers. The input consists in sequences of consecutive words stripped of their diacritic symbols. The network's task is to learn to restore the diacritics for the entire sequence. No additional linguistic or semantic information is used as input to the networks. The best results were obtained with a CNN-based architecture and achieved an accuracy of 97% at word level. At diacritic-level the accuracy of the same architecture is 89%.

17:40 - 18:00
Source Code Retrieval for Bug Localization using Bug Report View Paper
Kyaw Ei Ei Swe Hnin Min Oo	University of Computer Studies, Mandalay, Myanmar University of Computer Studies, Mandalay, Myanmar
Abstract: Bug localization helps software developers to track post-released faulty source files with the help of user’s reported bug files. In recent, automatic information retrieval based bug localization tools are proposed. It recommends relevant faulty source files to repair according to their maximum similarity scores. The system proposes a combined approach of IR-based bug localization by operating previously fixed bug reports and structures of source code and bug report. From the query, bug report constructs are also considered to get more accurate bug localization. In the proposed system, the first similarity score is calculated by using the previously fixed bug. The second similarity score is calculated by structuring the source code files and bug report. In our approach, three parts in the source code file and two parts in bug report are combined as six combinations score. Finally, these two similarity scores are combined. In some combing bug localization approach, features are usually linearly combined normally. In the proposed system used linearly combine with the weight value. It performs experiments on three projects, i.e. SWT, AspectJ, and Eclipse. The proposed system performs the experiments in three evaluation metrics. The result shows that the proposed approach achieves the relevance accuracy for bug localization process. According to the evaluation result, the proposed approach with structures is more localized than no structured approach.

Facial Expression Understanding (Friday, September 6, 16:40 - 18:00)
Location: Beijing Room, 5^th Floor

Chair: Radu Timofte
Co-Chair: Diana Borza

ETH Zurich, Switzerland
Technical University of Cluj-Napoca, Romania

16:40 - 17:00
An initial study of feature extraction’s methods in facial expression recognition View Paper
Tudor Tolciu Sinziana Toma Cristian Matei Laura Diosan	Babes-Bolyai University, Romania Babes-Bolyai University, Romania Babes-Bolyai University, Romania Babes-Bolyai University, Romania
Abstract: Facial Expression Recognition is a sub-branch of Affective Computing that specializes in extracting human emotions and facial features from visual input (images and videos) and tagging them to specific emotion hierarchies. The difficulty of this task lies not only in the subjectivity of distinguishing between human emotional states but, also in the diversity of the human race and culture that influences how we humans perceive sentiments. This article aims to tackle this challenge through two distinct methods of image processing, meaning automatic, Artificial Intelligence (AI) driven feature extraction and manual feature extraction through classical approaches and compare the performance of each of them afterwards. Our conclusion is, that both methods yield noteworthy results, and each specializes in a different context.

17:00 - 17:20
Efficient convolutional neural network for apparent age prediction View Paper
Casian Miron Vasile Manta Radu Timofte Alexandru Pasarica Radu-Ion Ciucu	Technical University of Iasi, Romania Technical University of Iasi, Romania Technical University of Iasi, Romania ETH Zurich, Switzerland Technical University of Iasi, Romania University Politehnica of Bucharest, Romania
Abstract: This paper proposes an efficient convolutional neural network architecture for apparent age estimation from single face image capable of top performance without the use of a pretraining stage. The architecture consists of 9 convolution layers and 2 max pool layers and requires only 79k parameters. We also improve the results by applying a weighted class distribution which assures that overwhelmingly represented classes do not skew the prediction results. These are compared to other results in literature for both methods that use pretraining and those that do not. The proposed method achieves estimation errors comparable to the human reference and to existing methods while being characterized by efficiency at both training and testing, it employs orders of magnitude fewer parameters and much less training time than other state-of-the-art methods.

17:20 - 17:40
On the link between facial expressions and emotional states induced by exposition to multimedia contents View Paper
Sérgio Paiva Herman Gomes	Universidade Federal de Campina Grande, Brazil Universidade Federal de Campina Grande, Brazil
Abstract: The explosive growth of digital videos has created new challenges for computer science. While many advances on video indexing, retrieval and summarization based on general, subject-independent, objective descriptors have been made in the past years, research on the use of individual subjective preferences and affective states is at the forefront of research and poses great challenges. In this article, we study the relationship between emotional states reported by viewers and their facial physiological changes observed during the display of different video genres. A dataset of twenty videos was created from YouTube video sharing platform. During the exhibition of the videos, the viewer's facial activities have been recorded and analyzed by means of Action Units (AUs). After that, emotional states self-reported by the viewers were assigned to video shots. Labels were divided into four categories, defined according to a discrete version of Russel's Circumplex emotion model. Different machine learning models were trained to test the relationship between the measured facial features and the self-reported emotional categories. We obtained k-fold cross validation accuracies that were above chance for the best learned models. As a result of this study, we concluded that AUs can indeed be used as an valuable tool to estimate emotional categories during exposure to audiovisual stimuli, and, therefore, should be used in further studies that take advantage of those categories to devise personalized multimedia retrieval and summarization approaches.

17:40 - 18:00
Investigation of Automatic Video Summarization using Viewer's Physiological, Facial and Attentional Features View Paper
Sérgio Paiva Herman Gomes	Universidade Federal de Campina Grande, Brazil Universidade Federal de Campina Grande, Brazil
Abstract: Video summarization aims at the selection of a concise and representative set of keyframes or video segments that allows the identification of the video content. In either cases, traditional summarization techniques usually work by segmenting the video into shots, representing video frames as feature vectors of color, texture, audio, among other features, clustering frames with similar features and selecting most representative keyframes or segments, sometimes guided by a video-to-summary ratio target. The resulting summaries are typically subject independent and do not take into account specific viewer’s behavior. Instead of using intrinsic features extracted from the video for summarization, in this article we study whether personalized (subject dependent) video summaries can be obtained from physiological, facial, and attentional data captured from the viewers. More specifically, we study the relationship between personalized video summaries reported by viewers and their data captured during the display of different video genres. A dataset of fifteen videos was used in the experiments. During the exhibition of the videos, the viewer's physiological, facial, and attentional data were recorded, analyzed and synchronized. Several machine learning models were trained to test our hypothesis. We obtained k-fold cross validation accuracies that were above the chance for the best learned models. As a result of this study, we conclude that it is possible to train a learning machine that can produce customized summaries that are closer to user preferences compared to randomly produced summaries.

Special Session: Automated Driving 4 (Saturday, September 7, 09:00 - 10:20)
Location: Presidential Room, 5^th Floor

Chair: Mihai Pomarlan
Co-Chair: Tiberiu Marita

University of Bremen, Germany
Technical University of Cluj-Napoca, Romania

09:00 - 09:20
Exploration of Autonomous Mobile Robots through Challenging Outdoor Environments for Natural Plant Recognition Using Deep Neural Network View Paper
Masoud Fathi Kazerouni Nazeer T. Mohammed Saeed Klaus-Dieter Kuhnert	University of Siegen, Germany University of Siegen, Germany University of Siegen, Germany
Abstract: Modernization of living environments and human activities have severe effects on many parameters and factors such as climate change and global warming, an increase of incidence and the severity of wildfires, land surfaces and ice sheets, ecological imbalance, change of fertility of the soil, flow of energy, food security, etc. In addition, human modernization has had a negative impact on biodiversity and the natural environment. An integral component of modernization is agriculture that associates with the outdoor environment and relevant issues. Automation of agricultural activities contributes to reducing the dependency on human labor and the harmful effects on the natural environment. The correct identification of plants in outdoor environments has been neglected and many critical environmental and non-environmental limits and factors, such as weather conditions, time, viewpoint, lighting conditions (illuminations and light intensity), distance, etc., have not been considered in existing plant recognition systems and technologies. Hence, there is a demand to develop mobile and real-time systems for plant recognition in natural environments. This paper addresses these challenges and introduces the application of autonomous mobile robot and semi-robots for recognition of natural plant species in outdoor environments. Furthermore, the proposed system presents the use of employing low-cost cameras, such as iPhone 6s, Canon EOS 600D and Samsung Galaxy Note 4, for plant recognition system in real-time. The performance of the system has been tested with a number of experiments in different years, 2017 and 2018, and at different times of day, morning and evening. The proposed system is a combination of new technologies involving deep learning concepts and an autonomous field robot to carry out precise plant recognition in challenging environments. The final accuracy of the mobile real-time system is 84.1666%.

09:20 - 09:40
What are you doing with that thing: Getting Household Service Robots to Explain Themselves View Paper
Mihai Pomarlan	University of Bremen, Germany
Abstract: We describe a knowledge-based approach through which a household robot can explain its actions. The approach uses a model of situations, stereotypical scenarios that the robot and/or user may participate in. Situations are also characterized by the items participating in the situation and their affordances. We describe a natural language generation pipeline including content selection heuristics tailored to the household robotics domain, and illustrate our approach on table-setting tasks.

09:40 - 10:00
Improved 3D Perception based on color monocular camera for MAV exploiting image semantic segmentation View Paper
Andrei Baraian Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: In this paper, we propose an improved 3D perception for MAV based on color monocular camera. We exploit the semantic segmentation of scene color images to derive semantic and geometric constraints on the 2D keypoints and 3D reconstruction. The high-level components of the system are ego-motion estimation of the drone and sparse 3D reconstruction and segmentation of the explored environment. The ego-motion and 3D structure components are highly coupled and any significant improvement in either one greatly influences the other one. The proposed method aims at reducing the outliers that come from dynamic objects by avoiding extraction of keypoints from entities that could potentially exhibit dynamic behavior (cars, trucks, pedestrians, etc.) exploiting the semantic segmentation of the image. Also, we derive a set of semantic and geometric constraints which improve the 3D reconstruction. The resulting 3D point cloud is semantically segmented by transfering the semantic class of the originating points from which it was triangulated. Ultimately, all these improvements have a great impact on the speed-up and accuracy of the optimization method used, in our case windowed bundle adjustment.

10:00 - 10:20
Semantic Segmentation Learning for Autonomous UAVs using Simulators and Real Data View Paper
Bianca-Cerasela-Zelia Blaga Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Deep learning requires large amounts of data for training models. For the task of semantic segmentation, manual annotation is time consuming and difficult. With the recent advances in game engines, simulators have become more popular as they can generate ground truth data for multiple sensors. In this paper, we make a thorough survey on the most recent and popular simulators and synthetic datasets, exploring solutions for semantic segmentation on images taken from drones. We also propose an extension of the CARLA simulator by introducing an aerial camera. We evaluate a deep learning model trained on both synthetic and real data, and present a new dataset which comprises both.

Intelligent Agents (Saturday, September 7, 09:00 - 10:20)
Location: Venezia Room, 5^th Floor

Chair: Mariana Mocanu
Co-Chair: Adrian Groza

University Politehnica of Bucharest, Romania
Technical University of Cluj-Napoca, Romania

09:00 - 09:20
ARMAX: A mobile geospatial augmented reality platform for serious gaming View Paper
Alexandru Predescu Mariana Mocanu Ciprian Lupu	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania
Abstract: The AR concept can be considered as one of the key technologies in the context of Industry 4.0 as it can be applied in many domains to provide context-aware interaction with the environment. While the concept of mixed reality is not new, the technology can now be found in different shapes and sizes, each tailored for a particular use case. In this paper we aim at defining a broader context while developing a universal platform for Serious Gaming for different industries that benefits from the most advanced AR implementation combined with the ubiquity of mobile devices. Some of the main challenges related to the 3D registration and tracking of virtual objects are discussed and the results are presented as obtained from real world testing. The calibration method is validated by simulation and while not being the most sophisticated solution it shows the experimental work and testing that is required for the accurate alignment of the real world and the virtual objects on the AR. The solution is presented with a focus on gamification and user experience in the context of a crowdsensing platform that aims to increase collaboration and participation in a Smart City environment.

09:20 - 09:40
A mentalist agent for guessing characters View Paper
Adrian Groza Loredana Coroama	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: We consider the problem of finding the minimal sequence of questions needed to identify an unknown element from a set of cardinality M. This task is common meet in games such as Guess Who, 20 Questions or Akinator.Our scenario is to identify a person based on features extracted from an image. The assumption is that the user thinks at any person described on the DBpedia. We do not store previous expert knowledge or user profile. The sequence of questions is built based on heuristics that favor the most relevant features: information gain, gain ration, probabilistic entropy. As we deal with features that are automatically extracted from images, the data is noisy. We test the performance of the method using simulated dialogues between software agent and human agent.

09:40 - 10:00
Dishonest behavior dynamics in the presence of influential agents View Paper
Emil Bucur	Babeş-Bolyai University, Romania
Abstract: Influential agents are present in society since the beginning of the first forms of social organization. More or less people have been interacting with these influential agents, either directly or indirectly. In some situations, their presence has generated a development of dishonest behavior. Interaction between influential agents and other members of society has often resulted in their marginalization. This research aims to explain why influential agents tend to become isolated without being able to form clusters of individuals of the same strategy in their neighborhood. However, a dynamic balance can be established for indefinite periods of time, between honest and dishonest individuals. The results were obtained using evolutionary computational algorithms and aim to explain how the presence of influential agents influences the punishment mechanisms.

10:00 - 10:20
Finding the Spreaders in a Graph Database View Paper
Riham Abdel Kader	Beirut Arab University, Lebanon
Abstract: The use of graph databases has increased significantly in the last decades, especially for storing data in social networks (e.g. Facebook, Linkedin, etc.). An important question that has received considerable research is finding the group of spreaders or the influential nodes in a social network. These nodes are defined to be the set of nodes who collectively can reach all other nodes in the network diffusing and spreading information. These nodes can be used to drive the whole network to a certain opinion or idea, therefore finding these nodes can be of high interest to politicians and opinion makers and drivers. In this paper, we provide an algorithm that can identify the minimal set of spreaders in a database graph. The algorithm is based on first extracting the strongly connected components of the graph, and then using these to identify the set of bases of the directed graph. Each base can be used as the set of spreaders that will allow a piece of information to flow and reach all other users in the database.

Image Processing (Saturday, September 7, 09:00 - 10:20)
Location: Beijing Room, 5^th Floor

Chair: Anca Andreica
Co-Chair: Romulus Terebes

Babes-Bolyai University, Romania
Technical University of Cluj-Napoca, Romania

09:00 - 09:20
Asymmetric Generalized Gaussian Distribution Parameters Estimation based on Maximum Likelihood, Moments and Entropy View Paper
Nafaa Nacereddine Aicha Baya Goumeidane	Research Center in Industrial Technologies, Algiers, Algeria Research Center in Industrial Technologies, Algiers, Algeria
Abstract: In this paper, we address the problem of estimating the parameters of Asymmetric Generalized Gaussian Distribution (AGGD) using three estimation mehods, namely, Maximum Likelihood Estimation (MLE), Moment Matching Estimation (MME) and Entropy Matching Estimation (EME). For this purpose, these methods are applied on an unimodal histogram fitting of an image corrupted with AGGD noise. Experiments show that the effectiveness of each method comparatively to the other one depends on the variation range of the shape factor.

09:20 - 09:40
Evolved Cellular Automata for Edge Detection in Binary Images View Paper
Alina Enescu Anca Andreica Laura Diosan	Babes-Bolyai University, Romania Babes-Bolyai University, Romania Babes-Bolyai University, Romania
Abstract: One significant and complex image processing task that may be used as a step in various complex processes is the edge detection. The mechanism of detecting the edges is called edge detector. The scope of this paper is to propose an edge detector based on evolved Cellular Automata (CA), for binary images. It was proved in many cases that CA may be successfully applied in various image processing tasks because they have a number of advantages over the traditional methods of computations: simplicity of implementation, the complexity of behaviour, parallelisation, extensibility, scalability, robustness. In the scope of this paper, two Evolutionary Algorithms are used to evolve the CA' rule to detect edge points in binary images.

09:40 - 10:00
Using Convolutional Neural Network for ImageEnhancement on Mobile Devices View Paper
Mihai Despotovici Irina Mocanu Lucia Rusu	University Politehnica of Bucharest, Romania University Politehnica of Bucharest, Romania , Romania
Abstract: This paper presents an application used to automatically enhance an image captured by a camera of a mobile device. The proposed solution consists of applying 4 image enhancement algorithms: gamma correction, saturation correction, white and black level correction. In order to parameterize the algorithms, we proposed a solution based on a convolutional neural network. The network was ported on a mobile device, keeping in mind to minimize computation resources and battery consumption. The mobile platform used is a smart phone, which has an accelerator, useful for massive processing of data caused by the neural network and a graphical processor, used to apply the 4 algorithms.

10:00 - 10:20
An improved NLM filter with increased noise robustness and adaptive similarity function View Paper
Claudiu Bic Romulus Terebes	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The paper proposes an improved Non-Local Means (NLM) approach that employs a prefiltering step prior to the evaluation of the patch similarity function and uses a modified similarity function allowing a better and a finer control of the influence of the patches involved in the filtering process. Experimental results show that the improved method allows the obtention of better visual, and objective measures-quantified results than the original NLM formulation and recent developments in NLM filtering. The results are close to the ones obtained with state-of-the-art approaches like BM3D (Block Matching 3D)

Image Processing using Deep Learning (Saturday, September 7, 10:35 - 11:55)
Location: Presidential Room, 5^th Floor

Chair: Kristína Malinovská
Co-Chair: Anca Marginean

Czech Technical University in Prague, Czechia
Technical University of Cluj-Napoca, Romania

10:35 - 10:55
Detecting Wearable Objects via Transfer Learning View Paper
Svatopluk Kraus Pavel Krsek Kristína Malinovská Matúš Tuna	Czech Technical University in Prague, Czechia Czech Technical University in Prague, Czechia Czech Technical University in Prague, Czechia Comenius University in Bratislava, Slovakia
Abstract: Transfer learning is a well known technique to circumvent the problem of small datasets in deep machine learning. It has been successfully used in the field of camera surveillance image processing which suffers from poor data quality and quantity. We focused on the task of wearable object detection, namely distinguishing if a person is or is not wearing a backpack. We created new annotations for the DukeMTMC-attribute dataset to overcome the discrepancies among the attributes. We explored transfer learning with a frozen feature extractor as well as the model fine-tuning, which turned out to perform much better. In both setups we found that the Densenet161 is the best from tested architectures. Our best model achieved about 92% balanced accuracy on the testing set.

10:55 - 11:15
Vision Inspection of Bottle Caps in Drink Factories Using Convolutional Neural Networks View Paper
Mahdi Bahaghighat Fereshteh Abedini Misak S’hoyan Arthur-Jozsef Molnar	Raja University, Iran Amirkabir University of Technology, Iran National Polytechnic University of Armenia, Armenia Babes-Bolyai University, Romania
Abstract: Nowadays, the growth and development of artificial intelligence is more evident than any other time. Consequently, the development of machine vision systems in the industry has been studied extensively. The provision of a high-level quality control system based on vision inspection in production lines is an important issue. It not only can increase the efficiency, but also derives the necessary tools to gather information about technical errors and faults in the target system. In this paper, we focus on the vision inspection of bottle caps in drink factories using convolutional neural networks. There are different cap encapsulation statuses which are classified as Normal Cap, Unfixed Cap, and No Cap. According to these classes, we can make a decision between either acceptance or rejection of a given bottle. Achieved results show that our optimized fine-tuning method based on the VGG-19 as an End-to-End deep learning approach can lead to more than 99% accuracy.

11:15 - 11:35
EmotionSense: Real-time emotional feedback from the audience View Paper
Andrada Farcas Anca Marginean	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: With the explosion of automation impacting the whole world, understanding human behavior and interfacing the man with the machine has become a predominant subject. The abstract goal of this project is to expose in real-time meaningful and compelling information regarding the emotions of an audience. For this, there will be presented an application that enables a transmitter to monitor and better understand the emotional impact of the presentation, so that one can adapt the speech on the go, or take into consideration further improvements for better engaging the public’s attention.

11:35 - 11:55
Dot Matrix OCR for Bottle Validity Inspection View Paper
Mircea Paul Muresan Paul Szabo Sergiu Nedevschi	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Identifying expiration dates on water bottles for fast industrial processing of a large amount of products mainly affects water bottling factories, food warehouses and supermarkets. The impact of this problem is the distribution of expired bottles of water or their existence on the market. One of the key issues for automatic character readers from bottles is dotted text. Furthermore, the transparent and curved nature of the containing bottle in the presence of water increases the difficulty of the text extraction. An optical character recognition solution using a convolutional neural networks is proposed to solve this issue. The proposed solution segments the input image to extract the bottle, detects the text region of interest, then performs pre-processing operations and finally converts the characters extracted from the region of interest on the bottle in human-readable characters. The proposed solution has real time performance and it achieves high quality results on the evaluation set.

Intelligent Systems with Applications (Saturday, September 7, 10:35 - 11:55)
Location: Venezia Room, 5^th Floor

Chair: Corina Cimpanu
Co-Chair: Camelia Lemnaru

"Gheorghe Asachi" Technical University of Iasi, Romania
Technical University of Cluj-Napoca, Romania

10:35 - 10:55
Genetic Multiobjective Optimisation with Elite Insertion for EEG Feature Selection View Paper
Lavinia Ferariu Corina Cimpanu	"Gheorghe Asachi" Technical University of Iași, Romania "Gheorghe Asachi" Technical University of Iasi, Romania
Abstract: Embedded Feature Selection (FS) ensures the selection of few, relevant features, by directly re-designing the classifier for subsets of features. Naturally, this problem is formulated as a multi-objective optimization (MOO) addressing to the accuracy of the classifier and the parsimony of the feature vector. In MOOs, common ranking techniques use dominance analysis for providing a partial sorting of the solutions. Unfortunately, dominance analysis can also promote solutions less useful for the application, usually placed at the extremities of the fronts. In order to gradually guide the search towards the middle of the fronts, this paper proposes an adaptive ranking algorithm with insertion of elites (ARE), which could be integrated in any MOO genetic algorithm. ARE incorporates two new procedures proposed for labeling the preferred solutions and for inserting elites in the less populated areas, whenever a biased exploration is detected. The experimental investigations illustrate that GA with ARE offers better results than NSGAII, both for electroencephalogram (EEG) feature selection problem (which likely involves weakly conflicting objectives) and MOOs with strongly conflicting objectives.

10:55 - 11:15
A Robust Vehicle Trajectory Prediction System View Paper
Lucian Cristea Rodica Potolea	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The task of driving, seemingly simple for humans, can, in some situations, puzzle even the most advanced AI systems in existence today. As extensive research is ongoing in this direction within the scientific community, this paper introduces a deep-learning based system than can accurately and robustly predict a vehicle’s near future trajectory. It does so by using a specifically built convolutional neural network that identifies the road’s shape based on disparity images, which are fused in an original way with lane marking information extracted from grayscale images. The prediction system outputs the trajectory in the form of predicted car movements which are then transformed into a road-augmented projection and simple steering wheel / brake commands, for easy visualization and understanding. The system’s performance is proven by a dedicated metric, as well as a visual analysis.

11:15 - 11:35
Space Breakdown Method View Paper
Eugen-Richard Ardelean Alexander Stanciu Mihaela Dinsoreanu Camelia Lemnaru Rodica Potolea	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Overlapping clusters and different density clusters are recurrent phenomena of neuronal datasets, because of how neurons fire. We propose a clustering method that is able to identify clusters of arbitrary shapes, having different densities, and potentially overlapped. The Space Breakdown Method (SBM) divides the space into chunks of equal sizes. Based on the number of points inside the chunk, cluster centers are found and expanded. Even if we consider the particularities of neuronal data in designing the algorithm – not all data points need to be clustered, and the data space has a relatively low dimensionality – it can be applied successfully to other clustering problems as well. The experiments performed on benchmark synthetic data show that the proposed approach has similar or better results than two well-known clustering algorithms.

11:35 - 11:55
Voice Pathology Detection using an Incremental System Combining Possibilistic SVM and HMM View Paper
Rimah Amami	Ecole Nationale d'ingénieurs de Tunis, Tunisia
Abstract: The voice pathology detection using automatic classification systems is a useful way to diagnose voice diseases In this paper, we propose a novel tool to detect voice pathology based on an incremental possibilistic SVM-HMM method which can be applied to serval practical applications using non-stationary or a very large-scale data in purpose to reduce the memory issues faced during the storage of the kernel matrix. The proposed system includes the steps of using SVM to incrementally compute possibilitic probabilities and then they will be used by HMM in order to detect voice pathologies. We evaluated the proposed method on the task of the detection of voice pathologies using voices samples from the Massachusetts Eye and Ear Infirmary Voice and Speech Laboratory (MEEI) database. According to the detection rates obtained by our system, the performance sounds robust, efficient and speed applied to a task of voices pathology detection.

Intelligent Networking (Saturday, September 7, 10:35 - 11:55)
Location: Beijing Room, 5^th Floor

Chair: Piroska Haller

Co-Chair: Vasile Dadarlat

University of Medicine, Pharmacy, Sciences and Technology
of Targu Mures, Romania
Technical University of Cluj-Napoca, Romania

10:35 - 10:55
Using Side-Channels to Detect Abnormal Behavior in Industrial Control Systems View Paper
Roland Bolboaca Bela Genge Piroska Haller	University of Medicine, Pharmacy, Sciences and Technology of Targu Mures, Romania University of Medicine, Pharmacy, Sciences and Technology of Targu Mures, Romania University of Medicine, Pharmacy, Sciences and Technology of Targu Mures, Romania
Abstract: As demonstrated by the large number of related studies, the field of anomaly detection in the context of Industrial Control Systems (ICS) has reached a certain level of maturity. However, we believe that additional research is needed in order to explore the side-channel specific parameters exposed by regular operations within ICS. The term "side channel" is a term usually found in cryptanalysis, where information about the behavior of the cipher, that is, the non-functional information, is used to break the ciphers. In the context of anomaly detection, side-channels denote non-functional information that can be derived from the normal operation of the system in order to infer the actual system state. This paper presents an anomaly detection system that explores the periodicity in ICS communications, where particular application-level operations are triggered periodically. To this end, we leverage the periodicity of a security protocol that has been implemented as part of our prior work to secure communications in ICS. We measure the deviations in the execution of the protocol's different phases in order to detect abnormal events that are caused at different levels in the architecture of the ICS. The main advantage of the developed approach is that it is protocol, software and application agnostic, making it suitable for legacy ICS as well. Experimental results are conducted in the context of a real industrial control system operating in a Romanian gas transportation network.

10:55 - 11:15
Pollution Probes Application: the impact of using PVDM messages in VANET infrastructures for environmental monitoring View Paper
Bogdan Iancu Istvan Illyes Adrian Peculea Vasile Dadarlat	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: An Ad-hoc Vehicular Network (VANET) is usually employed to provide road and traffic safety. Nevertheless, several other important applications could be developed, using VANET as an overlaying infrastructure. A prototype system for using VANETs as a tool to monitor environmental quality parameters in various areas (urban and rural areas), using vehicles as sampling instruments, is proposed in the paper. The data collected by the vehicles will then be forwarded, using the VANET infrastructure, to Environmental Protection Agencies for storing, analysis, visualization and further examination. The paper also evaluates how the use of Probe Vehicle Data Messages (PVDM) – used by the proposed system, would affect the delivery of high priority messages (for example road hazards, traffic safety), in this case DENM messages and proposes various constrains for PVDMs.

11:15 - 11:35
SMURD+ Medical Software View Paper
Gabriel Cristian Dragomir-Loga Marius Pop	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: It is sad to say but nowadays the accidents occur more and more often. When talking about life, we mostly talk about time, because when somebody is involved in an accident a countdown between life and death starts. The proposed solution is characterized by the following factors: the standardization principle (according to icd10 standard), integration of message queues and topics, establishing a proper model for messages, model which facilitates the internal memory processing in order to minimize the response time, including an optimistic validation-based protocol that guarantees the transaction consistency. The application tends to have a certain flow starting from introducing certain information about the victim`s status and concluding by being redirected to the closest medical unit that can treat the patient`s wounds, in order to facilitate the victim`s injuries treatment in the quickest time possible. Moreover, the way in which data is sent back to the ambulance and displayed is a very accessible one, so no longer time will be wasted by interpreting the results that were sent back from the SMURD+ Medical Software integrated system to the ambulance.

11:35 - 11:55
Forecasting Financial Markets using Deep Learning View Paper
Razvan Zanc Tudor Cioara Ionut Anghel	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Forecasting the behavior of financial markets represents an area of interest for many traders and investors due to the potential increase of capital which an accurate forecast can provide. The main objective of this paper is to predict the market behavior using Deep Learning techniques. We propose a stacked LSTM (Long Short-Term Memory) architecture on which we conducted several experiments on cryptocurrency and forex datasets. Our study reveals that in the context of financial markets, a high accuracy of a forecasted asset does not imply that the forecasted value will contribute positively to a profitable trading system.

Medical Imaging with Deep Learning (Saturday, September 7, 12:10 - 13:50)
Location: Presidential Room, 5^th Floor

Chair: Radu Badea

Co-Chair: Delia Mitrea

"Iuliu Hatieganu" University of Medicine and Pharmacy,
Cluj-Napoca, Romania
Technical University of Cluj-Napoca, Romania

12:10 - 12:30
Atrous Separable Convolutions for Heart Chamber Identification View Paper
Cristian Gabriel Turtoi	Babes-Bolyai University, Romania
Abstract: Semi-automatic and automatic segmentation of heart chamber and great vessel from 3D cardiac magnetic resonance images (MRIs) play a vital role in medical imagery. However, this task is challenging due to the hard to distinguish cardiac boundaries and low resolution images that were available. Due to the last advancements of deep learning techniques, we propose a method based on current state-of-the-art Encoder-Decoder Atrous Separable Convolutions model referred as DeepLabV3+, to automatically segment the myocardium and blood pool from 3D cardiac MRIs. We evaluated our method on HVSMR 2016 challenge dataset and the results demonstrate the robustness of the current approach, being competitive with the best method proposed in this challenge.

12:30 - 12:50
Towards Balancing the Complexity of Convolutional Neural Network with the Role of Optical Coherence Tomography in Retinal Conditions View Paper
Anca Marginean Adrian Groza Simona Delia Nicoara George Muntean Radu Razvan Slavescu Ioan Alfred Letia	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: Convolutional neural networks have shown impressive performance in the medical image domain, but medical experts are somewhat skeptical in their predictions since the features are not directly graspable. We are looking into one of the technical challenges, namely the explainability of the results, to try to find out some demonstration of the regions deemed abnormal by deep learning. Following the trend of heatmaps indicating which local morphology changes would modify the predictions, we are trying to verify the facilitation of the clinical understanding in the eyes of the ophthalmologist.

12:50 - 13:10
Hepatocellular Carcinoma Segmentation within Ultrasound Images using Convolutional Neural Networks View Paper
Flaviu Vancea Delia Mitrea Sergiu Nedevschi Magda Rotaru Horia Stefanescu Radu Badea	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania ”Prof. Dr. Octavian Fodor” Regional Institute of Gastroenterology-Hepatology, Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania
Abstract: Hepatocellular carcinoma (HCC) is the most common malignant liver tumour. As establishing a disease is difficult even for trained medical personnel, the need for automated tools to assist in diagnosis has increased. Currently, the golden standard for HCC diagnosis is the needle biopsy, this being an invasive, dangerous method. We develop computerized, noninvasive techniques, based on ultrasound images, in order to automatically detect the HCC tumour. In this paper, we exploit the power of deep learning for the purpose of segmenting HCC within ultrasound images. Multiple deep learning methods, based on Convolutional Neural Networks (CNN), were compared for this purpose. As this segmentation problem is a difficult one, due to the complexity of the ultrasound images, but also to the high class imbalance and to the low number of available training data, a study concerning the effect of multiple loss functions used in the training phase was performed, as well.

13:10 - 13:30
Hepatocellular Carcinoma Recognition in Ultrasound Images Using Textural Descriptors and Classical Machine Learning View Paper
Raluca Brehar Delia Mitrea Sergiu Nedevschi Monica Lupsor Platon Magda Rotaru Radu Badea	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania "Iuliu Hatieganu" University of Medicine and Pharmacy, Cluj-Napoca, Romania
Abstract: The classification of the liver tissue based on the analysis of the ultrasound images is an important task in the field of computer aided diagnosis, the ultrasonography providing a non-invasive, low-cost, imaging solution. An early, accurate diagnosis of the malignant tumours, based on those images, would be very useful for both patients and doctors. Within ultrasound images, the malignant tumours can be hardly distinguished by the human eye, the golden standard for diagnosis being the biopsy, an invasive, dangerous method. In our research, we develop computerized, non-invasive methods, based on advanced image analysis and recognition techniques, for performing tumour detection within ultrasound images. This paper proposes a method for hepatocellular carcinoma recognition in ultrasound images, by employing textural features obtained from the Gray Level Co-occurence Matrix (GLCM), combined with Local Binary Patterns (LBP). At the end, a true positive rate of about 72% was obtained for HCC, using ensembles of Adaptive Boosted classifiers.

13:30 - 13:50
Skin Lesion Diagnosis Using Deep Learning View Paper
Horea-Bogdan Muresan	Babes-Bolyai University, Romania
Abstract: Providing reliable computer aid to experts in medical fields is a very important objective as it can improve the quality of healthcare, as well as reduce the associated costs. Most models used in medical diagnosis use out of the box models, pretrained on the ImageNet dataset. In this paper we introduce a deep learning model based on the Inception-ResNet network. The model is capable of classifying skin lesions from dermoscopic images with good accuracy (83.96%), surpassing previously obtained results of 81.33% and 78%. We analyze the effectiveness of a few augmentation methods for increasing the performance of the model. We also discuss the results obtained in our experiments and we present several potential ways that can be explored in order to increase the precision, recall and accuracy of the model.

Deep Learning Techniques (Saturday, September 7, 12:10 - 13:30)
Location: Venezia Room, 5^th Floor

Chair: Gabriela Czibula
Co-Chair: Ion Giosan

Babeș-Bolyai University, Romania
Technical University of Cluj-Napoca, Romania

12:10 - 12:30
Generating Data using Monte Carlo Dropout View Paper
Kristian Miok Dong Nguyen-Doan Daniela Zaharie Marko Robnik-Sikonja	West University of Timisoara, Romania West University of Timisoara, Romania West University of Timisoara, Romania University of Ljubljana, Slovenia
Abstract: For many analytical problems the challenge is to handle huge amounts of available data. However, there are data science application areas where collecting information is difficult and costly, e.g., in the study of geological phenomena, rare diseases, faults in complex systems, insurance frauds, etc. In many such cases, generators of synthetic data with the same statistical and predictive properties as the actual data allow efficient simulations and development of tools and applications. In this work, we propose the incorporation of Monte Carlo Dropout method within Autoencoder (MCD-AE) and Variational Autoencoder (MCD-VAE) as efficient generators of synthetic data sets. As the Variational Autoencoder (VAE) is one of the most popular generator techniques, we explore its similarities and differences to the proposed methods. We compare the generated data sets with the original data based on statistical properties, structural similarity, and predictive similarity. The results obtained show a strong similarity between the results of VAE, MCD-VAE and MCD-AE; however, the proposed methods are faster and can generate values similar to specific selected initial instances.

12:30 - 12:50
Complete Visualisation, Network Modeling and Training, Web Based Tool, for the Yolo Deep Neural Network Model in the Darknet Framework View Paper
Roxana Mihaescu Eduard Barnoviciu Serban Carata Mihai Chindea Marian Ghenescu Veta Ghenescu	UTI Grup, Romania UTI Grup, Romania UTI Grup, Romania UTI Grup, Romania Institute of Space Science, Romania Institute of Space Science, Romania
Abstract: This paper presents an interface designed for the Darknet neural network, used by the state of the art YOLO (You Only Look Once) models. The object detection is still representing a challenge in the field of computer vision, and this interface has the main purpose of providing a way to generate more easily new networks, in order to obtain the desired object detection system. Furthermore, through this interface, it can be generated and displayed the feature maps at any point in the network. Therefore, the Darknet interface presented in this paper, leads to a better understanding of how a neural network makes a certain decision regarding the final predictions. And so, newer and better algorithms can be developed, in order to solve the object detection problem.

12:50 - 13:10
A method for the measurement and interpretation of neuronal interactions: improved fitting of cross-correlation histograms using 1D-Gabor Functions. View Paper
Ana Maria Ichim Adriana Nagy-Dabacan Raul C. Muresan	Technical University of Cluj-Napoca, Romania Transylvanian Institute of Neuroscience, Romania Transylvanian Institute of Neuroscience, Romania Transylvanian Institute of Neuroscience, Romania
Abstract: Cross-correlation analysis of separable multi-unit activity is the most used method to investigate neuronal connectivity. Features such as peaks, troughs, and satellite peaks in the cross-correlogram reflect the temporal relation between the activities of neurons. Precise estimation of such features requires independent measures. A very popular and effective method is to perform curve fitting using 1D Gabor functions. However, because of the non-linearity of the function, an iterative fitting procedure using optimization algorithms is required. As proposed from literature, we used the Levenberg-Marquardt algorithm. However, when applied to our data, the algorithm performed poorly. Here, we show that Trust Region algorithm represent a more attractive alternative to Levenberg-Marquardt in terms of performance and computational cost.

13:10 - 13:30
Analyzing Meteorological Data Using Unsupervised Learning Techniques View Paper
Andrei Mihai Gabriela Czibula Eugen Mihuleț	Babeș-Bolyai University, Romania Babeș-Bolyai University, Romania National Meteorological Administration, Romania
Abstract: Weather nowcasting which is the analysis and short-term weather forecast is a topic of major interest both for the meteorological and machine learning researchers. The problem is a complex one due to the large volume of data (such as radar, satellite or other ground meteorological observations) which has to be analyzed by meteorologists for issuing nowcasting warnings. In addition, climate changes are chaotic and the climate models are complex. The main goal of the paper is to better understand the relationships between the meteorological products extracted from radar observations both in severe and normal weather conditions. Self organizing maps are proposed as unsupervised learning models for detecting hidden patterns in radar data which are well correlated with weather changes. Experiments performed on real radar data provided by the Romanian National Meteorological Administration highlight the potential of unsupervised learning to uncover in radar data hidden rules which are relevant from a meteorological perspective. The results of our study suggest promising results in applying predictive supervised learning models for weather nowcasting based on radar data.

Distributed Systems 2 (Saturday, September 7, 12:10 - 13:30)
Location: Beijing Room, 5^th Floor

Chair: Raul C. Muresan
Co-Chair: Eneia Nicolae Todoran

Transylvanian Institute of Neuroscience, Romania
Technical University of Cluj-Napoca, Romania

12:10 - 12:30
Detecting Non-Redundant Collective Activity of Neurons View Paper
Medorian Gheorghiu Adriana Nagy-Dabacan Raul C. Muresan	Transylvanian Institute of Neuroscience, Romania Transylvanian Institute of Neuroscience, Romania Transylvanian Institute of Neuroscience, Romania
Abstract: Brain activity is characterized by the activity of huge number of neurons that are densely connected in complex networks. Detecting how neural activity is coordinated across time and space over brain circuits is a very challenging task. In particular, ensembles of neurons may co-activate in specific patterns whose expression correlates to various stimuli, cognitive states, or behavioral outcomes. These patterns are expressed on a variety of timescales, from milliseconds to hundreds of milliseconds or seconds. Here, we extend a recently-introduced method for extracting stereotypical firing patterns by applying several post-processing steps that enable the precise estimation of pattern wavefronts irrespective of the integration timescale used to detect patterns. Using electrophysiology data recorded in mouse visual cortex, we show that this approach enables a more precise estimation of the relation between firing patterns and meso-scale dynamics, reflected in the local-field potential. The method removes the redundancy and blurring generated by the convolution of spike-trains with exponentially-decaying kernels and offers a sharper representation of the expression of firing patterns in brain data.

12:30 - 12:50
Experiences in building a distributed Earth Observation Platform View Paper
Marian Neagul Silviu Panica Teodora Selea	West University of Timisoara, Romania West University of Timisoara, Romania West University of Timisoara, Romania
Abstract: In this article we present a Proof-Of-Concept distributed system designed for supporting a Regional Earth Observation Platform, particularly the use-cases from the ESA Pathfinder EO4SEE Project. The described system is designed adhering to the ESA Thematic Exploitation Platform Architecture and is built using state of the art cloud technologies, aiming to provide a flexible, easy extensible and adaptive platform suitable for current Earth Observation applications but also supporting new, emerging, technologies.

12:50 - 13:10
Continuation-Based Metric Semantics for Concurrency View Paper
Eneia Nicolae Todoran	Technical University of Cluj-Napoca, Romania
Abstract: By using the mathematical methodology of metric semantics and continuation semantics for concurrency (CSC) we design semantic models for a concurrent language extended with multiparty synchronization based on Hoare’s CSP. We present a new kind of domain for CSC, where the structure of continuations is expressed by using functions which map computations to computations. We study the abstractness of CSC, by using a new optimality criterion specific to continuation semantics, that we call κ weak abstractness - a weaker version of Milner’s full abstractness criterion.

13:10 - 13:30
A Flexible Windows Workspace Saving and Restoring Utility View Paper
Bálint Szabó Adrian Coleșa	Technical University of Cluj-Napoca, Romania Technical University of Cluj-Napoca, Romania
Abstract: The aim of this paper is to present an implemented solution to the problem of saving and restoring the state of a workspace (a collection of running user applications and the characteristics of their opened windows). The motivation for this application came from the need of a tool capable of saving and restoring the state of a set of applications. The existing solutions have some limitations, in the sense that they are not exactly tailored for the use case of saving and restoring a workspace exactly as it looked at the time of saving: current solutions do not try to restore an application to the same screen location where the application was located when the scanning was done. With this knowledge we proceeded with our plan to create solution that restores the context of an application (to a certain degree) and places that application at its original screen position. When scanning the workspace we iterate through all the windows created by user applications. For each window, we record the window position, size, and then we query some properties of the application which owns the respective window: the executable path and the command line with which the application was launched. Later, using this information we can restore an application and thus a collection of applications (a workspace) to the original state. In this paper we present our method of solving this problem coupled with the design and implementation of an open-source software that puts our solution into action.

Thursday, September 5, 2019

Friday, September 6, 2019

Saturday, September 7, 2019

Detailed Technical Program

Workshops

Deep Learning in Perception and Control (Thursday, September 5, 14:00 - 15:50) Location: Beijing Room, 5th floor

Bosch Student Workshop In Automated Driving (Thursday, September 5, 16:10 - 18:00) Location: Beijing Room, 5th floor

Keynote Lectures

Plenary Presentation 1 (Friday, September 6, 09:00 - 09:50) Location: Ballroom, 1st floor

Plenary Presentation 2 (Friday, September 6, 10:10 - 11:00) Location: Ballroom, 1st floor

Plenary Presentation 3 (Friday, September 6, 11:20 - 12:10) Location: Ballroom, 1st floor

Conference Sessions

Special Session: Automated Driving 1 (Friday, September 6, 13:30 - 14:50) Location: Ballroom , 1st Floor

Data Driven Robotic Intelligent Systems Workshop (Friday, September 6, 13:30 - 14:50) Location: Venezia Room, 5th Floor

Special Session: HiPerGRID (Friday, September 6, 13:30 - 14:50) Location: Beijing Room, 5th Floor

Special Session: Automated Driving 2 (Friday, September 6, 15:05 - 16:25) Location: Ballroom , 1st Floor

Knowledge and Ontology Engineering (Friday, September 6, 15:05 - 16:25) Location: Venezia Room, 5th Floor

Distributed Systems 1 (Friday, September 6, 15:05 - 16:25) Location: Beijing Room, 5th Floor

Special Session: Automated Driving 3 (Friday, September 6, 16:40 - 18:00) Location: Ballroom , 1st Floor

Natural Language Processing (Friday, September 6, 16:40 - 18:00) Location: Venezia Room, 5th Floor

Facial Expression Understanding (Friday, September 6, 16:40 - 18:00) Location: Beijing Room, 5th Floor

Special Session: Automated Driving 4 (Saturday, September 7, 09:00 - 10:20) Location: Presidential Room, 5th Floor

Intelligent Agents (Saturday, September 7, 09:00 - 10:20) Location: Venezia Room, 5th Floor

Image Processing (Saturday, September 7, 09:00 - 10:20) Location: Beijing Room, 5th Floor

Image Processing using Deep Learning (Saturday, September 7, 10:35 - 11:55) Location: Presidential Room, 5th Floor

Intelligent Systems with Applications (Saturday, September 7, 10:35 - 11:55) Location: Venezia Room, 5th Floor

Intelligent Networking (Saturday, September 7, 10:35 - 11:55) Location: Beijing Room, 5th Floor

Medical Imaging with Deep Learning (Saturday, September 7, 12:10 - 13:50) Location: Presidential Room, 5th Floor

Deep Learning Techniques (Saturday, September 7, 12:10 - 13:30) Location: Venezia Room, 5th Floor

Distributed Systems 2 (Saturday, September 7, 12:10 - 13:30) Location: Beijing Room, 5th Floor

Deep Learning in Perception and Control (Thursday, September 5, 14:00 - 15:50)
Location: Beijing Room, 5^th floor

Bosch Student Workshop In Automated Driving (Thursday, September 5, 16:10 - 18:00)
Location: Beijing Room, 5^th floor

Plenary Presentation 1 (Friday, September 6, 09:00 - 09:50)
Location: Ballroom, 1^st floor

Plenary Presentation 2 (Friday, September 6, 10:10 - 11:00)
Location: Ballroom, 1^st floor

Plenary Presentation 3 (Friday, September 6, 11:20 - 12:10)
Location: Ballroom, 1^st floor

Special Session: Automated Driving 1 (Friday, September 6, 13:30 - 14:50)
Location: Ballroom , 1^st Floor

Data Driven Robotic Intelligent Systems Workshop (Friday, September 6, 13:30 - 14:50)
Location: Venezia Room, 5^th Floor

Special Session: HiPerGRID (Friday, September 6, 13:30 - 14:50)
Location: Beijing Room, 5^th Floor

Special Session: Automated Driving 2 (Friday, September 6, 15:05 - 16:25)
Location: Ballroom , 1^st Floor

Knowledge and Ontology Engineering (Friday, September 6, 15:05 - 16:25)
Location: Venezia Room, 5^th Floor

Distributed Systems 1 (Friday, September 6, 15:05 - 16:25)
Location: Beijing Room, 5^th Floor

Special Session: Automated Driving 3 (Friday, September 6, 16:40 - 18:00)
Location: Ballroom , 1^st Floor

Natural Language Processing (Friday, September 6, 16:40 - 18:00)
Location: Venezia Room, 5^th Floor

Facial Expression Understanding (Friday, September 6, 16:40 - 18:00)
Location: Beijing Room, 5^th Floor

Special Session: Automated Driving 4 (Saturday, September 7, 09:00 - 10:20)
Location: Presidential Room, 5^th Floor

Intelligent Agents (Saturday, September 7, 09:00 - 10:20)
Location: Venezia Room, 5^th Floor

Image Processing (Saturday, September 7, 09:00 - 10:20)
Location: Beijing Room, 5^th Floor

Image Processing using Deep Learning (Saturday, September 7, 10:35 - 11:55)
Location: Presidential Room, 5^th Floor

Intelligent Systems with Applications (Saturday, September 7, 10:35 - 11:55)
Location: Venezia Room, 5^th Floor

Intelligent Networking (Saturday, September 7, 10:35 - 11:55)
Location: Beijing Room, 5^th Floor

Medical Imaging with Deep Learning (Saturday, September 7, 12:10 - 13:50)
Location: Presidential Room, 5^th Floor

Deep Learning Techniques (Saturday, September 7, 12:10 - 13:30)
Location: Venezia Room, 5^th Floor

Distributed Systems 2 (Saturday, September 7, 12:10 - 13:30)
Location: Beijing Room, 5^th Floor