Technical Program

Thursday, October 28, 2021

In the program below, click on the session you want to join and you will be directed to the Teams meeting.
The start time of the presentations should be maintained as specified in the technical program, hence in case of a no show we recommend the session chairs to use the slot for discussions and questions related to previously presented papers.

*The schedule uses EEST (Bucharest) time zone !

	Track 02
14:00-15:50	Workshop Bosch Students Workshop
15:50-16:00	Coffee Break
16:00-17:20	Workshop Semantic and Geometric Visual Perception (TUCN)
15:50-16:00	Coffee Break
17:30-18:50	Workshop on Big Data and Machine Learning in CloudUT (TUCN)
18:50-19:50	Registration

Friday, October 29, 2021

8:00: Registration

	Track 01	Track 02
08:50-09:00	Welcome & Opening
09:00-09:50	Plenary Presentation: Alexandru Paul Condurache
09:50-10:00	Coffee Break
10:00-11:20	Special Session: HiPerGRID	Automated Driving 1
11:20-11:30	Coffee Break
11:30-12:50	Special Session: Cloud Computing	Automated Driving 2
12:50-14:00	Lunch
14:00-15:20	Intelligent Distributed Computing and Networking	Machine Learning Based Image Processing
15:20-15:30	Coffee Break
15:30-16:50	Intelligent Systems 1	Computer Vision 1
16:50-17:00	Coffee Break
17:00-17:50	Plenary Presentation: Mihai Surdeanu
17:50-18:00	Coffee Break
18:00-19:20	Intelligent Systems 2	Computer Vision 2

Saturday, October 30, 2021

8:00: Registration

	Track 01	Track 02
09:00-10:20	Natural Language Processing	Computer Vision Applications 1
10:20-10:30	Coffee Break
10:30-11:50	Machine Learning 1	Computer Vision Applications 2
11:50-12:00	Coffee Break
12:00-13:20	Machine Learning 2	Intelligent Signal Processing

Detailed Technical Program

Workshops

Bosch Student Workshop (Thursday, October 28, 14:00 - 15:50)
Teams Track 2

Workshop Details: here.

Semantic and Geometric Visual Perception (TUCN) (Thursday, October 28, 16:10 - 17:20)
Teams Track 2

Workshop Details: here.

Workshop on Big Data and Machine Learning in CloudUT (TUCN) (Thursday, October 28, 17:30 - 18:50)
Teams Track 2

Workshop Details: here.

Keynote Lectures

Plenary Presentation 1 (Friday, October 29, 09:00 - 09:50)
Teams Track 1

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

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

A Safety View on Generalization for Machine Learning
Dr. Alexandru Paul Condurache	Robert Bosch GmbH

Plenary Presentation 2 (Friday, October 29, 17:00 - 17:50)
Teams Track 1

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

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

Explainable Deep Learning for Natural Language Processing
Dr. Mihai Surdeanu	The University of Arizona

Special Session: HiPerGRID (Friday, October 29, 10:00 - 11:20)
Teams Track 1

Chair: Nicolae Tapus
Co-Chair: Ciprian Oprisa

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

10:00 - 10:20
Asteroid Image Classification Using Convolutional Neural Networks View Paper
Cosmin Rosu Victor Bacu
Abstract: An impact of an asteroid on planet Earth could have catastrophic consequences. To eliminate this threat, the first step that needs to be done is to identify in advance all the asteroids near Earth. In this paper, we present a convolutional neural network model which is able to classify astronomical images according to the presence of asteroids. The model is trained with real data from Isaac Newton Telescope (INT) located in La Palma, Canary Islands, Spain. The goal of this work is to create an automated system for asteroids detection and to reduce as much as possible the false negative rate. We show that the model reaches a 94% recall when it is trained with individual images.

10:20 - 10:40
Fast Clustering for Massive Collections of Malicious URLs View Paper
Robert Rentea Ciprian Oprișa
Abstract: A large number of network security attacks use the HTTP protocol and rely on malicious URLs. Security solutions can filter such URLs using pattern-based rules, but may also trigger false alerts. Generally, the blocked URLs need to be further analyzed in order to tweak the rules and eliminate the false positives. This analysis is hardened by the large amount of detected URLs. We propose a clustering algorithm that groups together similar URLs, to avoid analyzing small variations of the same data. Further, we propose a method for selecting some representative centroids from each cluster, such that each URL in the cluster is similar with at least a centroid. By employing this method the amount of URLs that need to be analyzed is greatly reduced. The experimental evaluation showed that the algorithm produces quality clusters with a reasonably small running time.

10:40 - 11:00
Anomaly Detection in Software Defined Wireless Sensor Networks Using Recurrent Neural Networks View Paper
Vlad Lazar Sorin Buzura Bogdan Iancu Vasile Dadarlat
Abstract: Software-defined wireless sensor networks (SDWSN) are emerging as efficient mechanisms to orchestrate sensor networks at large scale using central network controllers - high-performance devices with a global view of the entire network. In the proposed approach, an anomaly is an erroneous or unwanted behavior in the sensors' plane. SDWSNs, with their complex infrastructure, can benefit from anomaly detection to improve several features of the system, such as security or Quality of Service (QoS). The computation power available at the controllers can be leveraged to allow training a recurrent neural network model (RNN) to detect sensors exhibiting anomalous behavior. The current work proposes a new anomaly detection solution based on the autoencoder principle and LongShort Term Memory (LSTM) recurrent neural networks, using a deep-learning approach. Multiple simulations were run using real sensor data and the accuracy of the detected anomalies was measured. The results show that the RNNs can detect both short-term and long-term anomalies with a 87% accuracy. Thus, SDWSNs and RNN can be used as a first step in successfully detecting various types of anomalies.

11:00 - 11:20
bhyve - JSON format and capsicum support for the snapshot feature View Paper
Ionut Mihalache Maria-Elena Mihailescu Darius Mihai Mihai Carabas Nicolae Tapus
Abstract: The current implementation for the snapshot feature of the bhyve supervisor uses three files to save the state of the guest virtual machine. One of the files has binary format, which makes the debug process and data interpretation, two laborious processes. The file descriptors that are used for the files containing the state of the guest virtual machine have more permissions than needed. Without all the necessary restrictions bhyve, the BSD hypervisor, is vulnerable because an attacker can open new files, sockets or write to read only file descriptors.

Automated Driving 1 (Friday, October 29, 10:00 - 11:20)
Teams Track 2

Chair: Ivo Haering
Co-Chair: Tiberiu Marita

Fraunhofer EMI, Germany
Technical University of Cluj-Napoca, Romania

10:00 - 10:20
Framework for safety assessment of autonomous driving functions up to SAE level 5 by self-learning iteratively improving control loops between development, safety and field life cycle phases View Paper
Ivo Haering Florian Lüttner Andreas Frohrath Miriam Fehling-Kaschek Katharina Ross Armin Rupalla Frank Hantschel Christian Schyr Thomas Schamm Steffen Knopp Daniel Schmidt Andreas Schmidt Michael Frey Daniel Grimm Marc René Zofka Alexander Viehl Yang Ji Zhengxiong Yang Norbert Wiechowski
Abstract: Safety verification and validation of autonomous driving functions up to SAE level 5 pose enormous challenges for car manufacturers. The paper argues that efficient improvement opportunities arise by suitably combining iterative development and verification processes that use selflearning approaches and well-defined quality and convergence criteria within a conceptual framework. The following cycles are used: development cycle, safety life cycle and field life cycle. For these cycles, suitable phases are first identified and defined. Then linkages are given that enable criteria-based iterative execution and improvement of selected combined phases. For this purpose, the selected phases are further resolved. It is distinguished between local loops within one cycle and loops between several cycles as well as with respect to the time horizon they cover. Suitable sample machine learning (ML) and artificial intelligence (AI) methods for the improvement loops are proposed in order to improve safety assessment of autonomous driving (AD) functions. The article presents three different types of ML/AI approaches regarding their usage within the development process of AD functions as well as identifies further improvement potentials. The approach is illustrated by ML/AI approach examples for the efficient provision of relevant and critical scenarios for the training and assessment of AD functions.

10:20 - 10:40
Pose Based Pedestrian Street Cross Action Recognition in Infrared Images View Paper
Raluca Didona Brehar Cristian Cosmin Vancea Mircea Paul Mureșan Sergiu Nedevschi Radu Dănescu
Abstract: In the context of a traffic scenario captured during night with infrared cameras we focus on pedestrian street cross action and we study the influence of the pedestrian pose with respect to the road environment on the accuracy of the action recognition model. This paper presents a complete framework that performs pedestrian cross action recognition for infrared sequences captured mainly during night but also during day time. The main contribution of the paper resides in the study of the variation in pedestrian action recognition accuracy provided by the combination of pedestrian pose-based features with several road context features given by semantic segmentation networks. The main modules of the proposed framework consist in a YOLO based infrared pedestrian detector combined with a tracking algorithm that enhances the detections. A CNN based pose estimator is applied on detected pedestrians in order to extract the relevant keypoints of the pedestrian skeleton. Several semantic segmentation networks like U-Net, FCN and PSPNet have been adapted in order to perform the semantic segmentation of the road in infrared images. Pose features are combined with road context features provided by the semantic segmentation and input to a LSTM based cross action recognition network. The obtained results provide a 90% accuracy on CROSSIR dataset.

10:40 - 11:00
Safety Assessment of Partially Automated Lane Change System Controller Calibration by Classifiers View Paper
Egemen Karabiyik Rıfat Uzun Batuhan Durukal Erhan Ozkaya
Abstract: In the development phase of autonomous driving, safety is the most crucial aspect that needs to be satisfied before stepping into the validation process in the real-world. To provide the safe reactions of autonomous features, calibration parameters need to be tuned carefully according to safety requirements. Since the virtual platforms are beneficial regarding cost, time, and safety manners; calibration parameters can be classified in terms of safe and unsafe states in that platforms for a predefined scenarios and environments which can represent the testing conditions. In this paper, we proposed a method that performs safety-based calibration assessment employing the machine learning algorithms to model the relation between calibration parameters of Partially Automated Lane Change System (PALS) and safety status of the various driving events. Utilizing this method, we have shown that calibration parameters can be limited with a range that satisfies the defined safety related key performance indicators (KPI) based on the relevant standards.

11:00 - 11:20
Novel SKIP Features for LIDAR Odometry and Mappings View Paper
Asad Ullah Khan Dario Lodi Rizzini
Abstract: This work proposes a refined feature extractor for the LiDAR Odometry and Mapping (LOAM) algorithm often rely on features extracted from point clouds. This paper proposes a novel detection algorithm SKIP-3D (SKeleton Interest Point) for extraction of features namely edges and planner patches from multi-layer LiDAR scan. SKIP-3D make use of the organization of LiDAR measurements to search for salient points in each layer through an iterative bottum-up procedure. In the process it removes the low curvature points to find edges and classifies the points from point clouds acquired from different view points are associated and used for their alignment. Experimental results showed that Fast LiDAR Odometry and Mapping (F-LOAM) based on SKIP-3D feature extractor performs at least better than original F-LOAM feature extractor.

Special Session: Cloud Computing (Friday, October 29, 11:30 - 12:50)
Teams Track 1

Chair: Florin Pop

Co-Chair: Gheorghe Sebestyen

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

11:30 - 11:50
Distributed Sensor Data Collection Using Mobile Clouds for Public Transportation View Paper
Anders Skretting Tor-Morten Grønli
Abstract: Today mobile solutions are expected to streamline all kinds of daily tasks from entertainment and communication to more specific tasks such as ordering food or buying a ticket for public transport. Social networking solutions might entail frequent image processing while mobility solutions might activate sensors in order to provide public transport operators with insights into travel patterns. All of these tasks lead to an elevation in resource expenditure which becomes a problem for everyday mobile users, especially in terms of battery life. This work proposes the usage of mobile clouds in mobility related solutions in order to offload tasks and share context specific data with nearby devices. The proposed solution is a mobile application capable of ranking nearby nodes based on their performance characteristics, and then request these nodes to participate in a shared sensor data collection according to computing principles of mobile clouds. The proposed solution can be used to both reduce energy consumption of devices and increase the amount of sensor data collected. We evaluate the solution based on both energy consumption and data loss. We measure and compare both battery drainage and data loss where the latter is evaluated by measuring sensor events sent versus received.

11:50 - 12:10
Optimizing Towards a Multi-Cloud Environment through Benchmarking Data Transfer Speeds in Amazon Web Services and Google Cloud View Paper
Vlad Bucur Liviu Miclea
Abstract: Performance of I/O operations and robustness in managing data is a cornerstone of any computing system. Cloud computing has enabled vast storage capacity. As the world of software engineering moves closer to a multi-cloud environment it becomes imperative to understand what the differences between cloud providers are at the most basic level of service offered: data transfer. This paper aims to provide an updated set of benchmarks, using commercially approved and scientifically accurate tools to generate a performance test suite that will highlight key differences, and similarities, between two of the largest commercial cloud providers available: Amazon and Google. The purpose of these tests is to show the viability of a multi-cloud environment specifically by highlighting differences in speed, computational power, software and hardware between cloud providers.

12:10 - 12:30
Bin Packing Scheduling Algorithm with Energy Constraints in Cloud Computing View Paper
Marius-Florinel Tudosoiu Florin Pop
Abstract: Cloud computing systems are the backbone of our technology needs in everyday life and are one of the major electric energy consumers globally. Any improvement that can be added to the energy efficiency of these vast systems constitutes a big gain worldwide in our never-ending battle with climate change and pollution. This paper proposes a new algorithm for task scheduling in Cloud systems based on the bin packing algorithm using a greedy implementation in conjunction with an optimization algorithm for resource task execution. By executing tasks in as much time as possible without impacting customer experience too much due to latency and by using resources at close to 100% in order to use the minimum number of servers we can achieve on average a 3.79% increase in energy efficiency. On top of that, our algorithm is robust to extreme variations of incoming task deadline distributions.

12:30 - 12:50
IoT data collection and analysis services on CloudUT View Paper
Anca Hangan Zoltan Czako Gheorghe Sebestyen
Abstract: As part of CloudUT, a cloud infrastructure and platform for research, we introduce the IoT service with the objective of creating a support infrastructure for research projects that involve sensor-based monitoring and data analysis tasks. This service aims to reduce the costs of acquiring data storage and processing resources and to reduce the development time and cost of custom software components. In this paper we present the functionalities provided through this service. Moreover, we describe the IoT service architecture, its components, as well as the challenge of re-designing and extending existing tools to obtain a cloud-native solution that can be scaled to accommodate multiple users and to reduce the execution time of the analysis procedures.

Automated Driving 2 (Friday, October 29, 11:30 - 12:50)
Teams Track 2

Chair: Frank Köster
Co-Chair: Florin Oniga

German Aerospace Center (DLR), Germany
Technical University of Cluj-Napoca, Romania

11:30 - 11:50
Free Space Detection from Lidar Data Based on Semantic Segmentation View Paper
István Nagy Florin Oniga
Abstract: In this paper, a free space detection approach is proposed for the detection of drivable area in the context of autonomous driving. This approach aims to detect free space using only 3D LIDAR data as input, by organizing the LIDAR measurements in their sensor specific layer/channel representation. The unstructured 3D point cloud is converted into 2D panoramic images containing 3D coordinates related features. This representation allows employing semantic segmentation convolutional neural networks (CNNs) for the detection of the free space. Using CNNs, these images are semantically segmented into two classes: road and non-road. The final output is the free space region, that can be used for potential driving assistance functions. The proposed approaches are evaluated on a manually annotated set built from the KITTI road benchmark.

11:50 - 12:10
MVGNet: 3D object detection using multi-volume grid representation in urban traffic scenarios View Paper
Selma Deac Flaviu Vancea Sergiu Nedevschi
Abstract: Recent advances in 3D object detection focus on combining multiple 3D data representations in order to leverage each 3D representation. However, this opens a new issue: increased need for hardware memory and computational resources which are unfortunately limited. In this paper we explore a new 3D data representation, the Multi-Volume Grid representation for detecting 3D objects in traffic scenes. The MVG structure efficiently maps free and occupied space from complex environments, where hanging and protruding objects are present. In order to integrate the MVG structure, we design MVGNet, a multi-branch network. We capture the essence of the MVG’s dynamic sized pillar structures by proposing two new input point descriptors and by using simple yet efficient spatial and channel based attention mechanisms throughout our network. We evaluate our method on KITTI dataset on three classes: pedestrian, cyclist, car. Our method achieves state of the art results with a significant improvement for the pedestrian and cyclist classes.

12:10 - 12:30
Intelligent Co-Simulation Framework for Cooperative Driving Functions View Paper
Viktor Lizenberg Mhd Redwan Alkurdi Ulrich Eberle Frank Köster
Abstract: Cooperative driving, a technology domain that allows for autonomous as well as manually driven vehicles to cooperatively coordinate their maneuvers with the aid of intervehicular communication, represents nowadays a highly active scientific topic for numerous research projects, notably such as German funded project IMAGinE. In the development process of the corresponding cooperative driving functions, a big challenge is posed by their extensive and complex testing as well as verification and validation procedures, which is reasoned by the vast amount of relevant scenarios needed to consider, even when using modern simulation-based methods. In the work at hand, we introduce our novel co-simulation framework, involving a coupling of traffic flow simulation with vehicle dynamics simulation, as well as an integrated machine learning classification module, which is able to detect, generate and evaluate test scenes and scenarios. As a result, with our approach, we achieve an intelligent way to test and to evaluate the cooperative driving functions practically solely on relevant test scenarios, organized in a systematical workflow with reasonable effort.

12:30 - 12:50
Eyes Detector Approach for Driving Monitoring System for Occluded Faces without using Facial Landmarks View Paper
Myriam Vaca-Recalde Pedro López Garcı́a Javier Echanobe Joshue Pérez
Abstract: The current health situation with the use of masks complicates the analysis of gaze and head direction in driver monitoring systems based on facial detection since landmarks are not working properly. Due to this issue, the need to solve occlusion problems using an alternative method to the current ones has increased. On the other hand, the deployment of these systems inside the vehicles must be carried out in the least intrusive way possible for the driver. This article presents an approach for driver distraction analysis based on the driver’s eyes without using landmarks applying Deep Learning methods, and the study of different parameters such as detection speed for the deployment of the best accuracy-speed method in an embedded platform. Different state-of-the-art and open source neural networks have been used and tuned to address our current problem. On the other hand, as is well known, training these models requires an enormous amount of data. In the case of gaze, there are very few data sets dedicated specifically to it. UnityEyes software has been used to create the training and test datasets for the system since it creates the necessary amount of data needed by the models easily.

Intelligent Distributed Computing and Networking (Friday, October 29, 14:00 - 15:20)
Teams Track 1

Chair: Mariana Mocanu
Co-Chair: Vasile Dadarlat

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

14:00 - 14:20
Agents-as-a-Service - a novel approach to on-premise digital twins View Paper
Sebastian Braun Youssef Mostafa Jessica Ulmer Chi-Tsun Cheng Steve Dowey Jörg Wollert
Abstract: Enterprises face the challenge of keeping up with the fast pacing rate of digitalization and Industry 4.0. Systems and products increase in complexity; thus, data and information must be tracked accurately and technologies utilized efficiently. Two of the key enablers of these developments are digital twins (DTs) and industrial agents (IAs). Current literature offers definitions, architectures, and views of DTs. However, those views are not unified, and frameworks for DTs and IAs technologies are missing. We propose a general basis for a DT architecture as well as a multi-agent systems (MASs) framework that combines the concepts of DTs and IAs and is manufacturing focussed. Our architectural view on DTs generalizes the reference architecture model industry 4.0 (RAMI 4.0) and shows how IAs integrate into DT concepts. The linkage to RAMI 4.0 aims towards standardizing the architectural concepts of DTs and bringing them closer to a unity that is not only compatible with well-defined Industry 4.0 terminology but also open to non-manufacturing related use cases. Both proposed frameworks are used to describe and explain how a MAS can be realized according to the introduced models.

14:20 - 14:40
A distributed approach for increasing coverage in crowdsensing applications with focus on urban exploration and water infrastructure View Paper
Alexandru Predescu Diana Arsene Mariana Mocanu Costin Chiru
Abstract: The first step towards a sustainable society is about raising environmental awareness. Supported by technological advancements and the ubiquity of mobile devices, mobile crowdsensing benefits from gamification. Both crowdsensing and gamification contribute to increasing the level of awareness and the participation of citizens in the community. In this paper, the proposed solution aims to increase the coverage in mobile crowdsensing for collaborative infrastructure monitoring by leveraging the power of gamification and urban exploration, while the distributed approach is evaluated in terms of the global outcomes and performance in a simulated environment. The proposed dynamic mapping system integrates nearby search and dynamic feature generation for achieving an optimal distribution of incentives in a real-life exploration game.

14:40 - 15:00
An Automatic Machine Learning Approach to Ultra-Wideband Real Time Positioning View Paper
Vlad Ratiu Mihai Alexandru Aionitoaie Emanuel Puschita Vasile-Teodor Dadarlat
Abstract: Ultra-Wideband positioning systems are being used more and more for tracking both people and objects in dynamic environments. One of the most accurate positioning strategies in this context is the use of a Time Difference of Arrival data acquisition mechanism coupled to a multilateration approach. An alternative to this method is based on replacing multilateration with machine learning. In order to determine the optimum machine learning algorithm from a set of multiple options automatic machine learning is a valid possibility. The project described in this paper aims to implement automatic machine learning through the use of an auxiliary component, the Training and Evaluation Engine. Finally, machine learning results are compared with multilateration results, in order to determine if the presented approach brings improvement to the state of the art.

15:00 - 15:20
A Lockable ERC20 Token for Peer to Peer Energy Trading View Paper
Liana Toderean Claudia Antal Marcel Antal Dan Mitrea Tudor Cioara Ionut Anghel Ioan Salomie
Abstract: In this paper, we address the digitization of physical assets using blockchain technology focusing on energy and peer-to-peer trading on decentralized energy markets. Because they are forward markets and operate on a day-ahead timeline, the energy transactions are settled only at the movement of energy delivery. Having the option of locking the energy tokens by a third-party escrow becomes a highly desirable feature. Thus, we define a Lockable ERC20 token that provides the option for an owner to lock some of its tokens using smart contracts. A time-lock and an escrow party account or smart contract can be specified allowing the tokens to be unlocked when certain business conditions are met. For validation purposes, we have considered a peer-to-peer energy trading scenario in which the Lockable ERC20 token was used to digitize the surplus of energy of prosumers. In this case, the energy tokens committed in blockchain transactions are successfully locked up until the actual delivery of energy, the settlement considering the monitored data of energy meters.

Machine Learning Based Image Processing (Friday, October 29, 14:00 - 15:20)
Teams Track 2

Chair: Zalán Bodó
Co-Chair: Raluca Brehar

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

14:00 - 14:20
Semi-supervised Learning in Camera Surveillance Image Classification View Paper
Matúš Tuna Kristína Malinovská Igor Farkaš Svatopluk Kraus Pavel Krsek
Abstract: Recognizing pedestrian attributes in camera surveillance images is a very hard problem, due to the lack of high-quality labeled data. In the field of deep learning the semi-supervised learning paradigm provides a possible answer to this problem. We propose a novel semi-supervised model that we call Binary Mean Teacher, tailored for binary classification task of detecting the presence of wearable objects. We train our model in a traditional scenario with a randomly initialized model, but we also explore fine-tuning a model pretrained on a large-scale image dataset. The performance of our model is compared to strong supervised baselines trained or fine-tuned using our dataset and the same augmentation strategy as in our model. We evaluate the impact of various augmentation strategies commonly used in deep learning on the performance of models in our binary classification task. Using only 1000 labeled training images, randomly initialized Binary Mean Teacher model achieves roughly 90% classification accuracy compared to 75% accuracy of randomly initialized supervised model that does not use any augmentations. When both Binary Mean Teacher and the supervised model are pretrained using the ImageNet dataset, and augmentations are used for both models, the Binary Mean Teacher achieves 92% accuracy compared to 90% accuracy of the supervised model.

14:20 - 14:40
Feature axes orthogonalization in semantic face editing View Paper
László Antal Zalán Bodó
Abstract: Human image synthesis is the technology that allows a computer program to create realistic photos of non-existing people. Though it is a relatively novel research topic that is mostly used to synthesize human faces, generating moving human figures is also possible using this method. At first, composing the believable and realistic images was a complex process. To achieve decent results, photo-realistic modelling, animating and mapping of the soft dynamics of the human body was required. Nowadays these methods are replaced by approaches based on machine learning and neural networks. Our system is able to create realistic images, consisting of three main components. The first component is a Generative Adversarial Network (GAN) that can generate a random face from a noise vector. Secondly, a convolutional neural network is responsible to recognize facial features on the input photos. Lastly, a regression model computes the correspondence between the input noise vector and output features of the generated face. Using a well-known face dataset, we report results applying the newly proposed model and we also analyze the accuracy and the plausibility of these results.

14:40 - 15:00
An Actor-Critic Approach to Neural Network Architecture Search for Facial Expressions Recognition View Paper
Sergiu Cosmin Nistor
Abstract: Among the many communication channels that people use are the facial expressions. Many emotions and opinions are communicated through the movements of the facial muscles. Though this is an essential and natural form of communication for people, computers find if highly challenging to deduce the emotions that correspond to each state of the face. We propose in this work, as a proof of concept, an automatic method of generating increasingly superior convolutional neural network architectures capable of recognizing which are the emotions communicated through each facial expression. Our method is described in detail and its efficiency demonstrated. We present our results, including the best architectures that we discovered. These architectures obtained good results on publicly available datasets while employing a very small number of learnable parameters and being fast at both inference and training time.

15:00 - 15:20
Detecting residues of cosmic events using residual neural network View Paper
Hrithika Dodia
Abstract: The detection of gravitational waves is considered to be one of the most magnificent discoveries of the century. Due to the high computational cost of matched filtering pipeline, there is a hunt for an alternative powerful system. I present the use of 1D residual neural network for detection of gravitational waves. Residual networks have transformed many fields like image classification, face recognition and object detection with their robust structure. With increase in sensitivity of LIGO detectors we expect many more sources of gravitational waves in the universe to be detected. However, deep learning networks are trained only once. When used for classification task, deep neural networks are trained to predict only a fixed number of classes. Therefore, when a new type of gravitational wave is to be detected, this turns out to be a drawback of deep learning. Shallow neural networks can be used to learn data with simple patterns but fail to give good results with increase in complexity of data. Remodelling the neural network with detection of each new type of GW is highly infeasible. In this letter, I also discuss ways to reduce the time required to adapt to such changes in detection of gravitational waves for deep learning methods. Primarily, I aim to create a custom residual neural network for 1-dimensional time series inputs, which can learn a ton of features from dataset without giving up on increasing the number of classes or increasing the complexity of data. I use two of the classes of binary coalesce signals (Binary Black Hole Merger and Binary Neutron Star Merger signals) detected by LIGO to check the performance of residual structure on gravitational waves detection.

Intelligent Systems 1 (Friday, October 29, 15:30 - 16:50)
Teams Track 1

Chair: Rodica Potolea
Co-Chair: Eneia Nicolae Todoran

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

15:30 - 15:50
Continuation Semantics for Interaction and Concurrency View Paper
Eneia Nicolae Todoran
Abstract: We present an operational semantics designed in continuation-passing style for an abstract concurrent language providing a general mechanism of multiparty interactions. Using the classic notion of strong bisimulation, we show that the basic laws of concurrent systems are satisfied in this semantics. By customizing the behavior of continuations, we obtain semantic models for formalisms based on Milner's CCS extended with constructions for multiparty interactions.

15:50 - 16:10
Augmented Van Emde Boas Tree for Connected Vehicles Traffic Modeling View Paper
Ioan Stan Bogdan-Paul Ungur Rodica Potolea
Abstract: Fast growing of the vehicular transportation and urbanization increases the traffic in urban areas and traffic congestion becomes a bottleneck in most of the cities in the world. Intelligent Transportation Systems can be used to improve the driving experience in urban areas by efficiently share information between vehicles and use infrastructure tools to balance the traffic and improve the flow. In order to do this, such systems require an efficient model to represent the vehicles and their real-time GPS position on the roads. In this paper we propose a traffic data model that is used by a cooperative route planning strategy in order to predict and avoid traffic congestion in urban areas. For the proposed model we implemented three novel range query data structures based on Van Emde Boas Tree in order to represent traffic data in a Vehicle to Cloud infrastructure: Basic Augmented Van Emde Boas Tree, Layered Van Emde Boas Tree and Partial Sums Van Emde Boas Tree. Our proposed data structures and their corresponding algorithms were evaluated by simulating traffic congestion through randomly generated routes in an urban area. To the best of our knowledge, from the experimental results, we found that Partial Sums Van Emde Boas Tree overcomes by 12% the performance of the state-of-the-art data structure that models traffic. This fact makes Partial Sums Van Emde Boas Tree a valuable traffic modeling data structure that can be used to predict and avoid traffic congestion in a Vehicle to Cloud infrastructure.

16:10 - 16:30
Design and Implementation of a Markings Assisted Guide Robot for Art Museums View Paper
Ana Rednic Septimiu Crișan Radu Gabriel Dănescu
Abstract: As nowadays people are attracted by technology in all domains and fields of activity, museums are in danger of gradually losing their visitors, unless they find ways to attract younger people with new technologies and new ways of presenting their heritage. One of the most appealing solutions is the employment of robot guides to complement or replace a human guide. This would increase the novelty factor of a museum along with mitigating the problem of lack of personnel or experience. Considering the actual state of guided tours made by museum staff, a robot that guides visitors would be of a real interest and would offer experienced assistance any time. The robot proposed and implemented in this paper is catered to visitors or groups of visitors that want to participate in a guided tour where they could get streamlined information about presented works of art.

16:30 - 16:50
Real-Time Cognitive Evaluation of Online Learners through Automatically Generated Questions View Paper
Ritu Gala Revathi Vijayaraghavan Valmik Nikam Arvind Kiwelekar
Abstract: With the increased adoption of E-learning platforms, keeping online learners engaged throughout a lesson is challenging. One approach to tackle this challenge is to probe learners periodically by asking questions. The paper presents an approach to generate questions from a given video lecture automatically. The generated questions are aimed to evaluate learners' lower-level cognitive abilities. The approach automatically extracts text from video lectures to generate wh-kinds of questions. When learners respond with an answer, the proposed approach further evaluates the response and provides feedback. Besides enhancing learner's engagement, this approach's main benefits are that it frees instructors from designing questions to check the comprehension of a topic. Thus, instructors can spend this time productively on other activities.

Computer Vision 1 (Friday, October 29, 15:30 - 16:50)
Teams Track 2

Chair: Radu Danescu
Co-Chair: Delia Mitrea

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

15:30 - 15:50
OccTransformers: Learning occupancy using attention View Paper
Bogdan Maxim Sergiu Nedevschi
Abstract: Recent years have seen significant progress in learning based approaches for 3D reconstruction. While different kind of representations have been tried, signed distance fields start to gain interest as they are easier to train, do not leave such a big memory footprint and yield competitive results, often beyond state-of-the-art. Self attention based Transformer models achieve state-of-the-art in different fields like sentence translation, image recognition and even multi-dimensional object classification by reformulating the problems as a sequence-to-sequence prediction. In this work we show that a transformer based architecture achieves state-of-the-art results even in the field of 3D reconstruction from different inputs(single image, point cloud) without the need of specialized encoders or additional fine-tuning. Our architecture is trained separately for the tasks of noisy point cloud shape completion and for single image 3D reconstruction on the popular ShapeNet dataset, achieving state-of-the-art by a significant margin on the Chamfer distance metric and competitive results on IoU and Normal Consistency. We belive that transformer based architectures will play a significant role in the following years in the field of 3D reconstruction.

15:50 - 16:10
Object detection using part based semantic segmentation View Paper
Razvan Itu Radu Danescu
Abstract: Monocular vision systems are increasingly popular in driving assistance applications as they are easy to set up and do not require precise calibration or synchronization. The downside of monocular vision is the lack of 3D information, which makes the task of identifying individual objects that are close together in the image space difficult. The lack of 3D information must be compensated by high accuracy classification of the image data. This paper proposes a novel way of detecting objects using fully convolutional neural networks followed by lightweight geometric based post processing. The fully convolutional neural network has four semantic segmentation outputs corresponding to quarters of individual objects. Therefore, each pixel of the input image will be classified as either belonging to a top left, a top right, a bottom left, or a bottom right region of a whole object. If the object is occluded and only a few of the four regions are visible, the component pixels will still be labeled correctly. Based on the multiple outputs of the neural network, the pixels are grouped into connected regions using a clustering algorithm aware of the relations between the object’s quarters. The accuracy of individual obstacle instances is similar to the accuracy of the results obtained from instance segmentation networks, while the demand of resources and the number of trainable parameters is significantly reduced.

16:10 - 16:30
Stereo and Mono Depth Estimation Fusion for an Improved and Fault Tolerant 3D Reconstruction View Paper
Mircea Paul Muresan Raul Marchis Sergiu Nedevschi Radu Danescu
Abstract: Depth estimation approaches are crucial for environment perception in applications like autonomous driving or driving assistance systems. Solutions using cameras have always been preferred to other depth estimation methods, due to low sensor prices and their ability to extract rich semantic information from the scene. Monocular depth estimation algorithms using CNNs may fail to reconstruct due to unknown geometric properties of certain objects or scenes, which may not be present during the training stage. Furthermore, stereo reconstruction methods, may also fail to reconstruct some regions for various other reasons, like repetitive surfaces, untextured areas or solar flares to name a few. To mitigate the reconstruction issues that may appear, in this paper we propose two refinement approaches that eliminate regions which are not correctly reconstructed. Moreover, we propose an original architecture for combining the mono and stereo results in order to obtain improved disparity maps. The proposed solution is designed to be fault tolerant such that if an image is not correctly acquired or is corrupted, the system is still able to reconstruct the environment. The proposed approach has been tested on the KITTI dataset in order to illustrate its performance.

16:30 - 16:50
Obstacle Facets Detection from 3D LiDAR Measurements: First results View Paper
Marius Dulău Florin Oniga
Abstract: In this paper, we present first results for an obstacle facets detection approach. The proposed approach is suitable for environment perception with LiDAR sensors. Obstacle points are identified and clustered into object instances, using a channel-based clustering method. The contour of each object instance is extracted and used for facet extraction. Using an iterative approach based on Random Sample Consensus, facets are extracted along each obstacle contour. Each processing stage is designed in order to lower the processing time. For the evaluation, we compare the proposed approach with an existing approach, using the KITTI benchmark dataset. The proposed approach has similar or better results for certain obstacle categories, while being several times faster.

Intelligent Systems 2 (Friday, October 29, 18:00 - 19:20)
Teams Track 1

Chair: Claudia Lavinia Ignat
Co-Chair: Radu Razvan Slavescu

Universite de Lorraine, CNRS, Inria, France
Technical University of Cluj-Napoca, Romania

18:00 - 18:20
Hardware implementation of Hopfield-like neural networks: Quantitative analysis of FPGA approach View Paper
Sergiu Zaporojan Viorel Carbune Radu Razvan Slavescu
Abstract: Recent progress in AI is largely attributed to the development of machine learning, especially in the algorithm and neural network models. On the other hand, hardware is on the critical path for the future of AI in the big data era. The rapid development of embedded intelligence for machine learning applications is causing the systems to grow more and more complex. FPGA-based solutions are emerging as the right choice for the implementation of these applications. Obviously, it is very important to understand the impact of architectural parameters on the performance and hardware resources utilization. This paper provides a rigorous analysis of FPGA implementation of Hopfield-like neural networks. The relationship between the hardware resources used to synthesize the data path and those used to provide network connections is discussed, as well as the distribution of these resources and how it depends on the variation in the architectural parameters of the network. The analysis presented in this paper is based on Intel/Altera Cyclone FPGA devices.

18:20 - 18:40
Towards Predicting Merge Conflicts in Software Developments Environments View Paper
Marina Bianca Trif Radu Razvan Slavescu
Abstract: Software development is nowadays a collaborative process. Although there exist version control systems such as Git, that help with the challenges that collaborative work poses, the process still has some issues, one of them are merge conflicts. When two or more developers need to combine the contents of different branches, merge conflicts might appear if the same lines of code have been changed in parallel. The process of deciding which version is to be kept requires human assistance and is time-consuming as well as error-prone. In order to mitigate the risk, a possible solution is finding conflicts before they appear by using Machine Learning (ML) / Deep Learning (DL) based predictors. We aim to explore which of these approaches would fit best to the task. To this end, we trained various classifiers using datasets composed of Git-related features regarding open source repositories. We studied over 30 thousand merge scenarios and over 30 different candidate features. Our classifiers’ values for precision and recall range between 0.68 and 0.78, with remarkably high values for ROC AUC. Cascading a Random Forest classifier with a deep neural network one offered the best results for the task at hand. The results indicate that conflict prediction is indeed attainable using machine learning, which could be an important step towards better synchronization in software development.

18:40 - 19:00
Linking Gamification Preferences to Personality Traits in Computer Science Education View Paper
Imre Zsigmond Horia F. Pop
Abstract: Gamification remains an often used but poorly understood field of study. Continued basic research of specific mechanics expands our understanding. Our previous experiment with instant feedback and storylines, in university-level computer science education, yielded positive results. We wanted to try to enhance the previous results while exploring psychological traits towards a personalized gamified experience. Specific changes yielded significant results on nearly all metrics, compared to the previous experiment. All the while we gathered big five psychological profile data, that we linked to gamification preferences for personalized gamification.

19:00 - 19:20
Users trust assessment based on their past behavior in large scale collaboration View Paper
Claudia-Lavinia Ignat Quang-Vinh Dang
Abstract: In a large scale peer-to-peer collaboration where control over data is given to users who can decide with whom to share their data, a main challenge is how to compute trust in the collaborators. In this paper we show how to automatically compute users trust according to their past behavior during the collaboration in order to be able to predict their future behavior. We focus on two use cases: contract-based multi-synchronous collaboration and trust game from game theory. We show that computing trust from a single interaction between two users depends on the application domain, but that a general methodology for aggregating trust during the successive interactions between two users can be applied for both use cases.

Computer Vision 2 (Friday, October 29, 18:00 - 19:20)
Teams Track 2

Chair: Matti Kutila
Co-Chair: Vlad-Cristian Miclea

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

18:00 - 18:20
A Method for Automatic Radar Azimuth Calibration using Stationary Targets View Paper
Alexandru Bobaru Corina Nafornita Cristian Vesa
Abstract: Radar is one of the fundamental sensors used in ADAS (Advanced Driver Assist System). Automotive related functionalities like ACC (Adaptive Cruise Control), LCA (Lane Change Alert), CTA (Cross Traffic Alert) or TA (Turn Assist) require very precise detection and ranging of traffic and environment, otherwise, the whole ADAS performance can be degraded. Vehicle integration and mounting tolerances will influence the angular performance of the radar sensor, due to its installation behind a bumper or a cover and even due to aging or exposure to accidents or vehicle vibration. In this paper, we introduce a real time autocalibration method of an azimuth angular interval to correct the environmental influences on the sensor.

18:20 - 18:40
A survey on the current state of the art on deep learning 3D reconstruction View Paper
Bogdan Maxim Sergiu Nedevschi
Abstract: Efficient 3D reconstruction from different kind of inputs is a long standing effort of computer vision. Recent advancements in the field of machine learning, specifically deep learning, have started an interest in studying how well these techniques apply to the 3D reconstruction problem. Current efforts employ two main research directions: techniques applied to a single object, trying to reconstruct the surface as closely as possible from different kind of inputs and techniques applied to scenes made from multiple objects, which deal with topological representations, color, illumination and resource consumption. With a plethora of applications in computer graphics and computer vision, the results given by the deep learning techniques start to gain a serious position in the field of 3D reconstruction, yet no survey exist on the recent advancements on these new techniques. This survey summarizes the recent trend and applications of the deep-learning 3D reconstruction methods. We focus on learnable reconstruction methods from inputs like single image, multi-image and point cloud, using different representations, such as voxels, meshes and signed distance fields. Starting by presenting the datasets used, we follow by showing the main deep learning methods using different representations, presenting advantages and disadvantages. The second half of this survey focuses on scene reconstructions and open problems. Finally we conclude with a discussion of the importance of the 3D reconstruction and its possible applications in different fields such as automotive and mixed reality.

18:40 - 19:00
WildUAV: Monocular UAV Dataset for Depth Estimation Tasks View Paper
Horațiu Florea Vlad-Cristian Miclea Sergiu Nedevschi
Abstract: Acquiring scene depth information remains a crucial step in most autonomous navigation applications, enabling advanced features such as obstacle avoidance and SLAM. In many situations, extracting this data from camera feeds is preferred to the alternative, active depth sensing hardware such as LiDARs. Like in many other fields, Deep Learning solutions for processing images and generating depth predictions have seen major improvements in recent years. In order to support further research of such techniques, we present a new dataset, WildUAV, consisting of high-resolution RGB imagery for which dense depth ground truth data has been generated based on 3D maps obtained through photogrammetry. Camera positioning information is also included, along with additional video sequences useful in self-supervised learning scenarios where ground truth data is not required. Unlike traditional, automotive datasets typically used for depth prediction tasks, ours is designed to support on-board applications for Unmanned Aerial Vehicles in unstructured, natural environments, which prove to be more challenging. We perform several experiments using supervised and self-supervised monocular depth estimation methods and discuss the results. Data links and additional details will be provided on the project’s Github repository.

19:00 - 19:20
LiDAR system benchmarking for VRU detection in heavy goods vehicle blind spots View Paper
Topi Miekkala Pasi Pyykönen Matti Kutila Arto Kyytinen
Abstract: This article is related to using modern LiDARs and neural networks based algorithms for vulnerable road users (VRU) detection. The problem is obvious especially when considering blind spots of heavy goods vehicles. LiDARs have developed a lot recently and the results indicate that adults can be detected up-to 75 m distance from the sensor even thought that pattern recognition requires sufficient point cloud-resolution. Two different LiDAR brands have been compared to understand cost-benefits between LiDAR technologies. The results have been conducted using an automated passenger car while considering feasibility to big trucks. Due to automotive requirements, the processing rate is also considered since usually, the main bottleneck is computation power, which is limited in automotive products. The used neural network algorithm is Yolo based and has been designed for VRU detection.

Natural Language Processing (Saturday, October 30, 09:00 - 10:20)
Teams Track 1

Chair: Florin Leon
Co-Chair: Anca Marginean

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

09:00 - 09:20
Using BERT for Multi-Label Multi-Language Web Page Classification View Paper
Codrut-Georgian Artene Marius Nicolae Tibeică Florin Leon
Abstract: With a very large and constantly growing number of web pages available on the Internet and considering the diversity of the topics that the content of web pages develops, automatic web page classification is becoming increasingly important. The text represents one of the main components of the content of web pages. In the last years, researchers have achieved state-of-the-art results in various natural language processing tasks, including text classification. This was possible mainly due to the development of language models that are trained on large text corpora and afterwards fine-tuned for specific tasks in a transfer learning manner. Such a model is Bidirectional Encoder Representations from Transformers (BERT), which has proven to be very effective for text classification. In this work, we propose a series of experiments to evaluate the effectiveness of the pre-trained multilingual BERT on multi-label multi-language web page classification. Overall, our proposed web page classifier achieves competitive results and one may conclude that it can be part of an automatic web page classification system.

09:20 - 09:40
Safer museum guide interaction during a pandemic and further. Using NLP in human interactive museum visits View Paper
Laura Ceuca Ana Rednic Emil Ștefan Chifu
Abstract: This paper introduces an NLP based solution for conversational agents in art exhibits. The conversational techniques (text-to-speech and speech-to-text) applied on chatbots are an important issue according to the user needs in different environments, art domains, and regions. The NLP agent is trained on relevant data by matching questions that are partially similar or contain only keywords from the knowledge base, assuring an understanding of a broader spectrum of questions. This paper proposes the theoretical background, the solution implementation, running scenarios, and experimental results as well. The application can be implemented in any kind of exhibition, as long as the knowledge base is adapted to the exhibits in focus.

09:40 - 10:00
Medical Question Entailment based on Textual Inference and Fine-tuned BioMed-RoBERTa View Paper
Andreea Maria Monea Anca Nicoleta Marginean
Abstract: Information understanding and retrieval have always been a challenge in natural language processing (NLP), especially in specific domains like the medical one. Natural language inference (NLI) is the task that focuses on understanding the information using inference, while recognizing question entailment (RQE) is used for question answering systems, establishing if two questions can have the same answer. NLI and RQE have gained popularity in the medical field with the increased language model performance. However, the number of approaches is still limited. This study uses the state-of-the-art language model, BioMed-RoBERTa, to be fine-tuned for NLI and RQE tasks on medical data. We propose a method to automatically build a dataset for question entailment(QE) based on a generalized definition of QE according to which there is an entailment between two questions iff there is an entailment between their answers. The built dataset was used for fine-tuning a model which obtains good results on MEDIQA 2019 RQE task.

10:00 - 10:20
Understanding Cooking Recipes’ Structure Using Grammars View Paper
Andreea Ilies Anca Nicoleta Marginean
Abstract: The main objective of this research is to understand the structure of cooking recipes using context free grammars, and to build corresponding parse trees for them, labeled with semantic roles. Grammars are used as to provide a white box alternative in a world dominated by transformers. The data is gathered from two recipe websites, containing 45 000 steps, made of a variable number of sentences. The solution uses the tools Lex and Yacc, in combination with features of AllenNLP. Identification of the symbols of the context free grammars is done with Lex for which the words were extracted and tagged as parts of speech using AllenNLP’s constituency parsing. The context free grammar is processed by Yacc and the idea guiding its description is that a part of speech can only take specific positions in a sentence. Meanwhile, prepositions and keywords help identify the proper semantic role for a group of words. The construction of the tree was developed in a bottom-up approach, employing C functions in the Yacc file. Therefore, the result is a portable JSON file, equivalent to a parse tree for a cooking recipe, where each group of words is labeled with an appropriate semantic role. The trees can be further utilized in the comparison and evaluation of cooking recipes. Additionally, this methodology can be applied in processing several types of instruction manuals.

Computer Vision Applications 1 (Saturday, October 30, 09:00 - 10:20)
Teams Track 2

Chair: Belhedi Wiem
Co-Chair: Ion Giosan

Altran Technologies, France
Technical University of Cluj-Napoca, Romania

09:00 - 09:20
Hardware and heterogeneous CNN for vision systems View Paper
Belhedi Wiem Kammoun Ahmed Hireche Chabha Hannachi Marwa
Abstract: With the increasing need for monitoring and control in daily life, artificial intelligence (AI) is gaining more autonomy to process massive data with minimal human intervention. Among AI algorithms, Convolutional Neural Networks (CNNs) are increasingly used as the core of many intelligent systems, including those that run on mobile and embedded devices. However, the CNN execution is computationally demanding especially on resource-limited. This makes the CNN use on embedded devices quite challenging. It is therefore natural to turn to hardware development to both speed up the CNN process, reduce the cost of resources and allow on-board processing. In this context, this paper proposes a full hardware implementation as well as an heterogeneous acceleration of the CNN process using Field Programmable Gate Arrays (FPGA). Both recognition accuracy and computational time were analysed for all the proposed implementation types that are namely: the software implementation (C and Python implementations), the full hardware implementation (VHDL), and the heterogeneous implementation (VHDL and C). From these results, we note that the hardware acceleration using FPGA is validated with a factor of 100 compared to the software implementation.

09:20 - 09:40
Metrics for Evaluating the Continuity Capabilities of Object Detection Systems View Paper
Calin Diaconu Cristina Pele Mihai Negru
Abstract: This paper proposes two novel metrics for evaluating the consistency of object detection systems on continuous video sequences. Its aim is to help differentiate between two such systems that have similar results if evaluated with traditional metrics, so two recomputed values for precision and recall are described. They are based on a continuity score, which received its name based on the fact that it favors detections that cover more grouped frames, rather than sparse sets of detections, for true positive detections. The opposite can be said about false positives, where fragmented ones are preferred. It relies on giving bonuses to detections that appear in consecutive frames, it allows a number of interruptions in such a series, before penalties are applied, and it is designed to be highly configurable. A procedure for verifying whether two different detections represent the same object is also described. Having this correlation method, the metrics can be used for simple object detection systems, but it can also be modified such that they are used with systems that assign ids to detected objects.

09:40 - 10:00
Visual Odometry Drift Reduction Based on LiDAR Point Clouds Alignment View Paper
Cătălin-Cosmin Golban Corvin-Petruț Cobârzan Sergiu Nedevschi
Abstract: Visual and LiDAR based odometry methods are key enablers for autonomous robots sub-problems such as mapping and localization or temporal aggregation and fusion of sensor data. When it comes to trajectory reconstruction the errors of independent local estimations accumulate over time, resulting in overall drifts from the real trajectory. This paper proposes a method for reducing the global drift by correcting the video-based frame-to-frame trajectory estimation at selected points in time based on the alignment of past and current LiDAR 3D measurements. The main strength of the proposed method is the formulation of the problem as a system of linear equations, which is an adequate numerical approximation as the corrective component is small enough. In this way we eliminate the need to deploy expensive non-linear optimization methods for calculating the correction transform.

10:00 - 10:20
Image features for vision-based robot manipulation based on deep reinforcement learning View Paper
Rui Li
Abstract: Deep Reinforcement Learning (DRL) provides a potential toolset that enables industrial robots to autonomously learn manipulation skills, but the learning efficiency (success rate within certain learning episodes) is the bottleneck. In this work, we ascertained well-designed environmental observations to be vital for improving efficiency. To determine the impacts of different observations, we conducted simulation experiments of robots grasping, and evaluated three popular categories of environment observations -positions of the Tool-Center-Point, raw images from a fixed viewpoint camera, and image features (Sobel, Laplacian, HOG, LBP). The results indicate "image features" proved to be superior to the others, they contribute to higher success rate and learning speed.

Machine Learning 1 (Saturday, October 30, 10:30 - 11:50)
Teams Track 1

Chair: Gabriela Czibula
Co-Chair: Adrian Groza

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

10:30 - 10:50
A machine learning approach for data protection in virtual reality therapy applications View Paper
Maria-Mădălina Mircea Rareş Boian Gabriela Czibula
Abstract: Health information is a protected asset that should be kept private. The tradability of personal data brings risks to the health information shared by users online. Authentication is a crucial first step when working to keep personal information private. Virtual Reality applications usually bypass application-specific authentication in favor of provider-specific authentication (e.g. Steam, etc). This approach is not ideal for health applications. Virtual Reality secure authentication can be difficult because most methodologies are not user-friendly. Previously proposed Virtual Reality authentication systems use PINs, Patterns, or 3D object sequences. We propose an authentication method based on dynamic movements (i.e. dance moves). Machine learning-based models are employed to determine if the received movement matches the user’s previously chosen movement. The proposed method thus balances security with a better versatility than the one provided by traditional methods.

10:50 - 11:10
Explainable Artificial Intelligence for Person Identification View Paper
Loredana Coroama Adrian Groza
Abstract: There is an increasing demand for explainable models in various applications. We interleave here explanations generated both from black-box models and white-box models. To achieve this we employ (i) a local model-agnostic technique (LIME), (ii) a theoretic game approach (SHAP), (iii) an example-based explanation technique (ExMatchina), and (iv) a model-specific method (KTrain). The running scenario is to identify persons from images and to explain the algorithmic decision. First, we extract features from images in order to train a white-box model and deploy explanation methods that work with tabular data. Second, we take advantage of transfer learning and fine-tuning to train black-box models and deploy image-based explanations methods. We aim to increase the interpretability of the system by providing explanations. We use the results to determine whether or not decisions are valid or if explanations are viable. Although many explanations techniques have been developed, there are no appropriate performance metrics to evaluate them. We propose two metrics to evaluate explanations quality. Based on these metrics, we demonstrate that LIME is unstable as it generates different explanations for same instance at multiple runs. Advantages and disadvantages of explanations techniques are also discussed and a user-grounded evaluation is performed. The evaluation study reveals that several explanation techniques were preferred by participants, followed by features listing, example-based explanations and pixels-based explanations.

11:10 - 11:30
Human Behavior and Anomaly Detection using Machine Learning and Wearable Sensors View Paper
Ioana Alexandra Bozdog Daniel-Nicusor Todea Marcel Antal Claudia Antal Tudor Cioara Ionut Anghel Ioan Salomie
Abstract: This paper addresses the problem of detecting and analyzing human behavior using a set of non-privacy invasive wearable sensors aiming to identify potential anomalies. This may be an important tool for increasing the independence and delaying the institutionalization of older adults allowing them to live alone in their homes with little support from caregivers. We propose an experimental web-based distributed system that incorporates data from wearable sensors and machine learning-based algorithms for monitoring the person's behavior and detection of anomalies. Various configurations of feature selection techniques and features as well as manual labeling for supervised learning have been used. In case of anomalies detected in older adult behavior, the caregiver is notified. Finally, we illustrate the system implementation and functionality considering Fitbit smart band sensor and integration with Fitbit Cloud. The results obtained using a public activity dataset with different configurations of machine learning anomaly detection algorithms and features are promising, showing an accuracy of 87% and an F1-score of 0.9.

11:30 - 11:50
Machine Learning-based Approach for Predicting Health Information Using Smartwatch Data View Paper
Viorica Rozina Chifu Cristina Bianca Pop Andrei Ciurianu Emil Ștefan Chifu Marcel Antal
Abstract: This paper presents a machine learning-based method for predicting health information in terms of heart rate, sleep quality, burned calories and performed physical activity for a person wearing a smartwatch. The proposed method consists of three main steps, namely, data pre-processing, feature selection, and training and testing. Based on a data set collected with a Fitbit Versa 2 smartwatch and by applying three different machine learning algorithms (i.e. Long Short-Term Memory, Multilayer Perceptron, and Back propagation from scratch) the method is able to predict the efficiency of the sleep, the number of burned calories, the average heart rate and the name of the activity performed by a person. To evaluate the experimental results the accuracy score and logarithmic loss metrics have been used.

Computer Vision Applications 2 (Saturday, October 30, 10:30 - 11:50)
Teams Track 2

Chair: Vasile-Ion Manta
Co-Chair: Razvan Itu

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

10:30 - 10:50
Compact Solution for Low Earth Orbit Surveillance View Paper
Radu Danescu Razvan Itu Mircea Paul Muresan Vlad Turcu
Abstract: Low Earth Orbit Objects (LEOs) are objects that circle our planet at a distance of less than 2000 km from the surface. Due to their small orbital radius, they move fast and are sometimes affected by atmospheric drag, meaning that their orbit will change in time. This orbit includes communication satellites, Earth observation satellites, but also space debris such as rocket bodies which will eventually reenter the atmosphere. The fast motion, the changing nature of the orbit, their sheer number, and the periodic reentry events, lead to the need of intense observation of their position. This paper presents a compact, portable system for surveillance of the LEO objects. The system is built with commercially available, low-cost items, and is capable of on-site acquisition and real time processing of images. The acquired images are processed by background subtraction, analysis of the difference between frames, extraction of elongated objects corresponding to the satellite streaks, and forming trajectories (tracklets) from consecutive detections. The emphasis on trajectories instead of individual object properties allows successful detection of faint objects, without a significant increase in false positives.

10:50 - 11:10
Pothole Detection for Visually Impaired Assistance View Paper
Stefan-Daniel Achirei Ioana-Ariana Opariuc Otilia Zvoristeanu Simona Caraiman Vasile-Ion Manta
Abstract: The global number of visually impaired is growing fast due to aging world population. People suffering of severe visual impairments face many difficulties in their daily routine. Pavement holes are a major problem for their navigation and walking. The proposed algorithm detects holes in the sidewalks and roads using a Convolutional Neural Network. Starting from the previously published research, this paper proposes a practical solution for pothole detection used in the Navigation Module of the Sound of Vision Lite (SoV Lite) project. YoloV5s, Mobilenet V1 and Mobilenet V2 Lite were trained on Nvidia RTX using the obtained dataset, then deployed for testing on Nvidia Jetson NX. Due to the mobile platform constraints Mobilenet V1 was chosen to be integrated in SoV Lite. Because objects which are closer have higher detection confidence but also because the visually impaired person wearing the system is interested in the dangers close to him, in practice, we limit the detections of negative obstacles within a range of 8m. By creating a region of interest the run time is also enhanced. For training and experiments we used in-house acquired frames using a ZED 2 Stereo Camera as well as publicly available data and annotated it for the specific task of detecting potholes and drains in the pavement and streets. The obtained dataset was augmented and made publicly available.

11:10 - 11:30
A Combinatorial Approach to Detection of Box Pallet Layouts View Paper
Ernesto Fontana Dario Lodi Rizzini Stefano Caselli
Abstract: This paper presents an algorithm for detection and pose estimation of cardboard parcel boxes in depth images based on combinatorial enumeration. It is designed for sensor-driven manipulation of pallets consisting of stacked planar layers chiefly using 3D range measurements. The proposed method initially detects the planar top layer of the pallet and its polygonal contour, possibly containing holes. Then, it enumerates the hypotheses about the layout of the pallet layer and estimates the best matching configuration. Experiments on a real dataset assess the feasibility of the proposed approach.

11:30 - 11:50
A Novel Multi-Model Approach to Real-Time Road Accident Prediction and Driving Behavior Analysis Using a Fully Connected Feed-Forward Deep Neural Network and a CoreML Object Detection Model View Paper
Diya Dinesh
Abstract: As the leading cause of death within the U.S., Road of research: accident frequency prediction [3,4,5], inclement accidents take over 38,000 lives every year. Efforts are being taken weather effect on accident patterns [6,7] and accident severity nationwide to reduce the accidents and fatalities. Previous studies prediction [5]. Studies have also been performed on effect of on the use of computer science for road safety were centered driver behavior on cause of accidents [8, 9]. However, the around analysis of historical data and prediction. This study outcomes and impacts of these studies are limited by one main proposes a novel solution, including real-time updates and factor, which is the lack of ability to predict these features to road safety, with the use of Artificial Intelligence and accidents/accident-prone zones and monitor/analyze driver Deep Learning integrated with various APIs and statistical behavior in real time. analyses through the RoadSafety application. This application consists of three features: accident risk prediction, landmark Real-time application is essential for drivers to get live analysis, and driving behavior analysis. The accident risk updates about how to stay safe while driving. The product of this prediction component consists of a fully connected feed-forward study (RoadSafety) is a solution to this limitation. Based on a deep neural network that takes in location, weather, time, and fully connected feed-forward deep neural network that analyzes road feature input to predict an accident risk level. The landmark location, road feature, time, and weather data, from the analysis identifies, through usage of the pearson correlation OpenWeatherMap API [27], RoadSafety produces live updates coefficient and recursive feature elimination, which types of on accident risk to drivers. This application also combines risk-locations/landmarks are best correlated with accident severity. prediction with statistical landmark analysis including the The driving behavior analysis uses an object detection CoreML pearson correlation coefficient and recursive feature elimination model and the pinhole projection formula to identify distance on data from GooglePlaces to warn users about locations and from the driver to an obstacle ahead. This feature also compares sites to be cautious around. In addition to this, RoadSafety the driver’s speed to the speed limit. All three features are monitors driver behavior in accident-prone zones by identifying integrated into an iOS application to provide drivers within D.C. obstacles with a CoreML model and calculating distance with live updates on accident prone-zones, landmark indicators of through the pinhole projection formula. It also surveys speed high accident severity, and risky driving behaviors.

Machine Learning 2 (Saturday, October 30, 12:00 - 13:20)
Teams Track 1

Chair: Camelia Lemnaru
Co-Chair: Mihai Negru

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

12:00 - 12:20
Underwater Noise Estimation with General Regression Neural Network View Paper
Belhedi Wiem François Rioult Achraf Drira Medjber Bouzidi
Abstract: With growing exploration and utilization of the ocean by human beings, sound pollution increases accordingly and significantly. Hence, the impact of noise on marine organisms has become one of the most important research topics. In order to assess the impact of these activities on marine fauna, and especially on marine mammals, underwater noise propagation should be estimated. It is in this context that the proposed work is situated. In fact, the aim of this project is to study how neural network techniques can speed up these computations for estimating underwater noise propagation. For this, experimental data from practical investigations/experiments were used to train the GRNN for estimating noise level caused by each boat. The predicted values using GRNN closely followed the experimental ones with a root mean square error scores (RMSE) that is equal to 0.7. Results showed that the GRNN model had good prediction results during the testing process in terms of both RMSE scores and training times. To evaluate the model accuracy, the propagation distance from the source in the horizontal plane was set at 150 km. At this distance, the predicted loss of acoustic energy resulted in noise levels which were comparable to the reference values. Several parameters was used in order to build an accurate model. These parameters include: the propagation distance, the depth, the transmission frequency, the kinematic bathymetry, and the sediment nature. The proposed model therefore estimated the noise level until 150 km from the noise source with high accuracy and high speed computation without complex procedures. Moreover, the use of GRNN made it possible to avoid remaking the expensive computations for each sub-zone: it estimated the noise levels with a very reduced time compared to the state-of-the art methods such as RAM.

12:20 - 12:40
Randomness Testing with Neural Networks View Paper
Imola Nagy Alin Suciu
Abstract: Testing the quality of data produced by Random and Pseudo-Random Number Generators (RNG-s and PRNG-s) is necessary for their safe use in cryptographic applications. Randomness is a probabilistic property and various statistical tests can be applied to evaluate the generated data sequences. This paper presents a novel approach for applying such tests, by training Artificial Neural Networks (ANN) to replicate the behaviour of both stand-alone and combined statistical tests. The process includes the development of various augmentation techniques used for creating a synthetic data set, the development and training of different ANN architectures and also the evaluation of the final classification models. The trained models can detect the presence or absence of diverse statistical characteristics in the generated data sequences. The proposed solution reaches over 0.95 accuracy for both stand-alone and combined application of the following statistical tests from the NIST Statistical Test Suit [1]: Frequency Test, Frequency Test within a Block, Runs Test, Tests for the Longest-Run-of-Ones in a Block.

12:40 - 13:00
Feature Selection via Genetic Multiobjective Optimization with Fuzzy Rejection Mechanisms View Paper
Corina Cîmpanu Lavinia Ferariu Tiberius Dumitriu
Abstract: In pattern identification, embedded Feature Selection (FS) determines a detailed data description and an efficient and accurate classification. Most Genetic Algorithm (GA) based optimization procedures tackle the minimization of a classifier's error rate. Independent of how feature selection procedures are configured, either in Single (SOO) or Multi-Objective Optimization (MOO) manners, the major problem with all classifiers is multidimensionality and quantity of redundant and noisy data recordings. This paper compares six GA-based optimizations as viable solutions for an accurate and efficient assessment of working memory load levels during arithmetic computations based on Electroencephalogram (EEG) data. An objective layout analysis of a randomly generated population motivates both SOO and MOO further explorations. The single objective error-based optimization, the aggregation of error, and the number of selected features illustrate the limitations of SOOs. The third SOO incorporates a fuzzy rejection mechanism for unnecessary features. The baseline for MOO comparison is Deb's NSGA. Two new MOO procedures for feature selection are proposed: a MOO Fuzzy Rejection of Irrelevant Features (MOO-FRIF) and a Fuzzy Progressive Deletion of Irrelevant Features (MOO-FPDIF). Both new MOO approaches gradually eliminate or discourage the EEG features that negatively influence the classifier. Two classifiers address the problem distinctively, namely Random Forests (RF) and a multi-class Support Vector Machine (SVM) with one vs. one comparison mechanism and RBF kernel.

13:00 - 13:20
End-to-End Automated Machine Learning System for Supervised Learning Problems View Paper
Mihai-Bogdan Bîndilă Mihai Negru
Abstract: Automated machine learning (AutoML) represents one of the fastest-growing fields in both industry and academia. Its popularity has increased over the last few years due to the rising demand for machine learning systems usable by non-experts. Existing open-source systems fail to fully automate the machine learning tasks, being more like a black box that generates non-reusable pipelines. We propose an end-to-end system accessible to non-specialists and data scientists who wish to effortlessly create reusable and easy to understand pipelines for any regression and classification problems. The proposed solution increases the automation level by performing an advanced data cleaning followed by a complete feature engineering process without any domain knowledge about the data set. These processes are integrated into a search that generates for each model the optimum instance of preprocessed data. Bayesian optimization (BO) with a Gaussian process (GP) maximizes the performance of these models that will be ensembled using standard and probability-based stacking. We prove that the system performs better than 43.47% of practitioners on real-world data sets, being robust to any variation of tabular data. The true potential is revealed by surpassing at least one similar tool in 50% of performed benchmarks. These results were achieved for any tested data set, regardless of its dimensionality, after a maximum of 30 minutes.

Intelligent Signal Processing (Saturday, October 30, 12:00 - 13:20)
Teams Track 2

Chair: Raul C. Muresan
Co-Chair: Mihaela Dinsoreanu

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

12:00 - 12:20
Weighted Principal Component Analysis based on statistical properties of features for Spike Sorting View Paper
Roxana Ioana Aldea Mihaela Dinsoreanu Rodica Potolea Camelia Lemnaru Raul Cristian Muresan Vasile Vlad Moca
Abstract: Spike Sorting is a challenging problem in Computational Neuroscience because of the complexity of neural data. One of the greatest issues are overlapping clusters. This paper focuses on the feature extraction step in the Spike Sorting pipeline and proposes an adaptation of Principal Component Analysis (PCA) to increase the separability between clusters. This is achieved by weighting the features before applying PCA, taking into consideration the multimodality and the distance between probability distributions. The information extracted from the characteristics of a multimodal distribution is the number of modes (peaks). The distance between the probability distributions is quantified using Jensen-Shannon divergence. The computed information, number of modes and distance, is aggregated into a coefficient representing the weight of the features. The new approach has been validated on a synthetic dataset and shows improvements compared with the state-of-the-art PCA.

12:20 - 12:40
Deep neural networks for prediction and detection of ocular sequelae among survivors of Stevens-Johnson syndrome/toxic epidermal necrolysis View Paper
Kevin Sheng-Kai Ma
Abstract: Ocular complications, including recurrent corneal inflammation, progressive corneal neovascularization, conjunctival epithelial ingrowth, and permanent vision loss, are the major long-lasting sequelae of Stevens-Johnson syndrome (SJS)/toxic epidermal necrolysis (TEN), a species of adverse drug reactions. As such, predicting the severity of ocular complications in chronic stage SJS/TEN allows for optimized ocular prognosis. In this study, with external eye photos, fluorescein staining, and photos of slit lamp examination during acute stage and chronic stage SJS/TEN, we first predicted ocular prognosis of chronic stage SJS/TEN using acute stage images, and then determined the severity of ocular complications of chronic stage SJS/TEN using chronic stage images, so as to facilitate the management of ocular complications of SJS/TEN with convolutional neural networks (CNNs). In our study, the prediction of chronic stage ocular complications of SJS/TEN during acute stage, and automatic detection of ocular complications of SJS/TEN during chronic stage, were both clinically satisfying when using external eye photos and photos of slit lamp examination for Schirmer’s test values, superficial punctate keratopathy (SPK), hyperemia, and Ocular Surface Grading Score (OSGS), but not for best corrected visual acuity (BCVA). Moreover, since the CNN model may accurately detect the age and sex of patients with the photos, it is suggested that our model possessed high validity. In conclusion, CNN-based multi-class classification may both predict ocular complications of SJS/TEN before their onset, and accurately detect them upon the development of these sequelae.

12:40 - 13:00
Jittered sampling - a potential solution for detecting high frequencies in GCaMP recordings View Paper
Andrei Ciuparu Raul Cristian Muresan
Abstract: Multi-photon microscopy is a widely used method to measure cortical activity through the use of calcium indicators called GCaMP. These methods have inherent limitations stemming from the fluorophore as well as from the microscope setup that effectively limit the rate at which the activity can be measured. According to the Nyquist-Shannon sampling theorem, this limits the bandwidth of activity that can be estimated using these methods. Here we introduce and test an extension to current recording setups based on irregular (jittered) sampling that may be able to partialle overcome these limitations. We generated synthetic datasets consisting of mixtures of oscillations at different frequencies, sampled far below the Nyquist rate. We show that by sampling the oscillations at multiple phases using a random delay between acquired samples it is possible to characterize the amplitude of an oscillation present in the signal. Furthermore, the predicted amplitude associated to aliases of the measured frequency is reduced. Finally, we show that it is possible to detect and characterize an oscillation packet embedded in a noisy, undersampled signal.

13:00 - 13:20
A Generative Adversarial Approach for the Detection of Typical and Drowned Action Potentials View Paper
Alexandru Dodon Marius Adrian Calugar Rodica Potolea Camelia Lemnaru Mihaela Dȋnşoreanu Vasile Vlad Moca Raul Cristian Mureşan
Abstract: Spike sorting methods are central for the interpretation of extracellular recordings and provide insight into more complex brain processes. To be able to apply the sorting methods spikes are detected, conventionally, by using an amplitude threshold that takes into account the features of the signal. This offers a straightforward detection solution that is able to discriminate the background noise, but comes at the cost of ignoring spikes with an amplitude below the threshold and limiting the picture provided by the recording. We propose a machine learning approach that captures spikes overlooked when considering only the amplitude. The proposed method leverages Generative Adversarial Networks (GANs) to learn particularities of the spike and noise components of the recordings and then extract the underlying components of novel samples with the goal of detecting the presence of spikes. We quantify the detection metrics with respect to the signal to noise ratio of the recording and show that the proposed method exhibits significantly higher sensitivity while providing comparable specificity to conventional spike detection methods across all signal environments, especially when spikes are almost drowned in the background noise.

ICCP 2021 Technical Program

Technical Program

Thursday, October 28, 2021

Friday, October 29, 2021

Saturday, October 30, 2021

Detailed Technical Program

Workshops

Bosch Student Workshop (Thursday, October 28, 14:00 - 15:50) Teams Track 2

Semantic and Geometric Visual Perception (TUCN) (Thursday, October 28, 16:10 - 17:20) Teams Track 2

Workshop on Big Data and Machine Learning in CloudUT (TUCN) (Thursday, October 28, 17:30 - 18:50) Teams Track 2

Keynote Lectures

Plenary Presentation 1 (Friday, October 29, 09:00 - 09:50) Teams Track 1

Plenary Presentation 2 (Friday, October 29, 17:00 - 17:50) Teams Track 1

Special Session: HiPerGRID (Friday, October 29, 10:00 - 11:20) Teams Track 1

Automated Driving 1 (Friday, October 29, 10:00 - 11:20) Teams Track 2

Special Session: Cloud Computing (Friday, October 29, 11:30 - 12:50) Teams Track 1

Automated Driving 2 (Friday, October 29, 11:30 - 12:50) Teams Track 2

Intelligent Distributed Computing and Networking (Friday, October 29, 14:00 - 15:20) Teams Track 1

Machine Learning Based Image Processing (Friday, October 29, 14:00 - 15:20) Teams Track 2

Intelligent Systems 1 (Friday, October 29, 15:30 - 16:50) Teams Track 1

Computer Vision 1 (Friday, October 29, 15:30 - 16:50) Teams Track 2

Intelligent Systems 2 (Friday, October 29, 18:00 - 19:20) Teams Track 1

Computer Vision 2 (Friday, October 29, 18:00 - 19:20) Teams Track 2

Natural Language Processing (Saturday, October 30, 09:00 - 10:20) Teams Track 1

Computer Vision Applications 1 (Saturday, October 30, 09:00 - 10:20) Teams Track 2

Machine Learning 1 (Saturday, October 30, 10:30 - 11:50) Teams Track 1

Computer Vision Applications 2 (Saturday, October 30, 10:30 - 11:50) Teams Track 2

Machine Learning 2 (Saturday, October 30, 12:00 - 13:20) Teams Track 1

Intelligent Signal Processing (Saturday, October 30, 12:00 - 13:20) Teams Track 2

Bosch Student Workshop (Thursday, October 28, 14:00 - 15:50)
Teams Track 2

Semantic and Geometric Visual Perception (TUCN) (Thursday, October 28, 16:10 - 17:20)
Teams Track 2

Workshop on Big Data and Machine Learning in CloudUT (TUCN) (Thursday, October 28, 17:30 - 18:50)
Teams Track 2

Plenary Presentation 1 (Friday, October 29, 09:00 - 09:50)
Teams Track 1

Plenary Presentation 2 (Friday, October 29, 17:00 - 17:50)
Teams Track 1

Special Session: HiPerGRID (Friday, October 29, 10:00 - 11:20)
Teams Track 1

Automated Driving 1 (Friday, October 29, 10:00 - 11:20)
Teams Track 2

Special Session: Cloud Computing (Friday, October 29, 11:30 - 12:50)
Teams Track 1

Automated Driving 2 (Friday, October 29, 11:30 - 12:50)
Teams Track 2

Intelligent Distributed Computing and Networking (Friday, October 29, 14:00 - 15:20)
Teams Track 1

Machine Learning Based Image Processing (Friday, October 29, 14:00 - 15:20)
Teams Track 2

Intelligent Systems 1 (Friday, October 29, 15:30 - 16:50)
Teams Track 1

Computer Vision 1 (Friday, October 29, 15:30 - 16:50)
Teams Track 2

Intelligent Systems 2 (Friday, October 29, 18:00 - 19:20)
Teams Track 1

Computer Vision 2 (Friday, October 29, 18:00 - 19:20)
Teams Track 2

Natural Language Processing (Saturday, October 30, 09:00 - 10:20)
Teams Track 1

Computer Vision Applications 1 (Saturday, October 30, 09:00 - 10:20)
Teams Track 2

Machine Learning 1 (Saturday, October 30, 10:30 - 11:50)
Teams Track 1

Computer Vision Applications 2 (Saturday, October 30, 10:30 - 11:50)
Teams Track 2

Machine Learning 2 (Saturday, October 30, 12:00 - 13:20)
Teams Track 1

Intelligent Signal Processing (Saturday, October 30, 12:00 - 13:20)
Teams Track 2