Thursday, September 3, 2020

Friday, September 4, 2020

Saturday, September 5, 2020

Detailed Technical Program 2020

Workshops

Bosch Student Workshop (Thursday, September 3, 14:00 - 15:50)
Channel: Track 03

Workshop Details: here.

Integrated Semantic Visual Perception and Control for Autonomous Systems (Thursday, September 3, 16:10 - 18:00)
Channel: Track 03

Workshop Details: here.

FROM 2020 (Thursday, September 3)
Channel: TBA

Workshop Details: here.

Keynote Lectures

Plenary Presentation 1 (Friday, September 4, 09:00 - 09:50)
Channel: Track Plenary

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

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

Camera Calibration
Knorr Moritz	Bosch Corporate Research, Germany

Plenary Presentation 2 (Friday, September 4, 14:00 - 14:50)
Channel: Track Plenary

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

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

Robotics and AI: Promises and Challenges
Daniela Rus	Massachusetts Institute of Technology, USA

Plenary Presentation 3 (Friday, September 4, 15:00 - 15:50)
Channel: Track Plenary

Chair: Sergiu Nedevschi
Co-Chair: Rodica Potolea

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

Building Knowledge For AI Agents With Reinforcement Learning
Doina Precup	McGill University and DeepMind, Canada

Conference Sessions

Computer Vision: Automated Driving 1 (Friday, September 4, 10:00 - 11:20)
Channel: Track 03

Chair: Matti Kutila
Co-Chair: Radu Danescu

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

10:00 - 10:20
Driver Assistance for Safe and Comfortable On-Ramp Merging Using Environment Models Extended through V2X Communication and Role-Based Behavior Predictions View Paper
Lucas Eiermann, Florian Wirthmüller, Kay Massow, Gabi Breuel and Ilja Radusch
Abstract: Modern driver assistance systems as well as autonomous vehicles take their decisions based on local maps of the environment. These maps include, for example, surrounding moving objects perceived by sensors as well as routes and navigation information. Current research in the field of environment mapping is concerned with two major challenges. The first one is the integration of information from different sources e. g. on-board sensors like radar, camera, ultrasound and lidar, offline map data or backend information. The second challenge comprises in finding an abstract representation of this aggregated information with suitable interfaces for different driving functions and traffic situations. To overcome these challenges, an extended environment model is a reasonable choice. In this paper, we show that role-based motion predictions in combination with v2x-extended environment models are able to contribute to increased traffic safety and driving comfort. Thus, we combine the mentioned research areas and show possible improvements, using the example of a threading process at a motorway access road. Furthermore, it is shown that already an average v2x equipment penetration of 80% can lead to a significant improvement of 0.33 [m/s^2] of the total acceleration and 12m more safety distance compared to non v2x-equipped vehicles during the threading process.

10:20 - 10:40
V2X supported automated driving in modern 4G networks View Paper
Pasi Pyykönen, Aki Lumiaho, Matti Kutila, Johan Scholliers and Geerd Kakes
Abstract: With connected and automated vehicles and remote-controlled driving, available 4G/5G network capacity performance is the key issue. Wide bandwidth allows vehicles to receive and send real-time sensor data to a MEC (Mobile edge computing) server or remote controller. Wide download bandwidth doesn’t ensure low latency with vital data, e.g. control commands or camera image. The main difference between 4G and 5G cellular networks is the capability to ensure wide bandwidth for uploading data from a vehicle to a MEC server but also low and predictable latency. This article studies baseline measurements conducted in the modern 4G network, taking into account the expected 5G features. This publication introduces the results from the VTT automated vehicle in the KPN (Dutch landline and mobile telecommunications company) cellular testnetwork at the automotive test-track between Helmond and Eindhoven, the Netherlands. The tests are carried out to understand potential performance improvements when updating 4G network to 5G network for autonomous vehicles. Measurements are taken by using two MQTT servers. The local MQTT server is the MEC server in the KPN network. The second reference server is located at VTT’s premises in Finland. This publication shows that 4G technology performance is not sufficient even in a dedicated test network with a controlled amount of simultaneous users.

10:40 - 11:00
A comparison study on replacing stereo disparity with LiDAR in visual odometry methods View Paper
Catalin Golban, Petrut Cobarzan and Sergiu Nedevschi
Abstract: The visual-based odometry estimation is a key feature in today's autonomous robots. Although the best performance for video-only solutions is achieved with stereo cameras, the LiDAR sensors’ general suitability for precise depth measurements at reasonable density, leads to the idea of fusing monocular cameras with LiDAR depth maps for the task of visual odometry. This paper presents a method for replacing depths computed using disparity estimation techniques with LiDAR based depth estimation, in the frame-to-frame stereo visual odometry method described in [2]. A comparative evaluation of the performance obtained by the same framework method using one or the other depth measurement source is presented in detail, concluding that the LiDAR measurements are suitable as input in existing visual odometry techniques, leading to increased estimation precision.

11:00 - 11:20
PartID – Individual Objects Tracking in Occupancy Grids Using Particle Identities View Paper
Radu Danescu, Razvan Itu and Mircea Paul Muresan
Abstract: Occupancy grid tracking algorithms see the world as made out of cells that can be either free or occupied, and can have speed probability densities attached to each cell. These algorithms estimate the overall state of the environment based on sensor data, but they are not aware of, nor concerned with the identity of individual objects. This paper proposes a new approach for individual objects tracking using the dynamic occupancy grids, which will embed the identity of the objects in the grid state. The particle based dynamic occupancy grid is extended by attaching identity information to each particle, thus achieving individual object tracking at grid level without the need of explicit modeling of the object’s shape. The position and dynamics of the world occupied cells are tracked independently of their identity, by the mechanism of the particle based occupancy grid. For achieving individual object tracking, this mechanism is extended with components for assigning and managing the identity of the particles. The designed system was tested on real world sequences, and was able to successfully track obstacles found on the road without making assumptions about their nature, shape or size.

Intelligent Systems: Fuzzy Reasoning and Brain Signal Classification (Friday, September 4, 10:00 - 11:20)
Channel: Track 01

Chair: Sergiu Zaporojan
Co-Chair: Mihaela Dinsoreanu

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

10:00 - 10:20
Integration of Fuzzy Logic in Analogical Reasoning: A Prototype View Paper
Moreno Colombo, Sara D'Onofrio and Edy Portmann
Abstract: Research and practice have long been working on the development and implementation of intelligent or smart systems. Such efforts are particularly made in the field of artificial intelligence. In this context, this article deals with the design and development of a system that should come close to human information processing. The proposed framework, consisting of an analogical scheme which is extended by fuzzy logic, represents an additional information processing module that can process and interpret unknown terms by linking them with known ones. The focus of this contribution is set on the creation of a prototype able of computing conceptual and spectral analogies. Thereby, this article discusses the current state of the fuzzy analogical reasoning prototype, and its implications and limitations.

10:20 - 10:40
Data-Based Technique for Extracting Knowledge from Data Generated in Experiments View Paper
Sergiu Zaporojan, Viorel Carbune and Igor Calmicov
Abstract: Fuzzy sets are used in different fields and determination of the membership functions is one of the most important issues in the design of fuzzy systems. The paper presents an approach to that problem to provide solutions in specific cases. In context, a technique for extracting knowledge from measurements data sets was developed that allows to retrieve human expertise and the construction of algorithms for decision-making systems. Initially, the method was developed to be used in data-based fuzzy modeling for the microwire casting plant control.

10:40 - 11:00
Using Symbolic Analysis of Local Field Potentials for Anesthesia Depth Prediction View Paper
Ioana Onofrei, Andreea Salagean, Narcisa Sırca, Vasile Moca, Adriana Nagy-Dabacan, Raul Muresan, Rodica Potolea, Camelia Lemnaru and Mihaela Dinsoreanu
Abstract: Depth of anesthesia (DOA) is a very important parameter, determining how the brain responds to various concentrations of anesthetic. Here, we assess the dependence of DOA on the concentration of isoflurane by analyzing local-field potentials (LFPs) recorded in mouse visual cortex. To capture the DOA-related changes, we employed TESPAR and Fourier-based feature extraction techniques, combined with machine learning. In addition, we performed an analysis of bursts present in LFP signals and evaluated their influence on DOA classification. These bursts are characterized by high amplitudes and low frequencies, requiring different analysis approaches: one that removes the high amplitudes and the other one that filters the lower frequencies. We developed procedures that successfully discriminate between different levels of anesthesia. Our study indicates a correlation between increased isoflurane dose and LFP amplitude suppression and also indicates that large amplitude bursts are expressed in the signal in a dose-dependent manner.

11:00 - 11:20
Enhancing the Classification of EEG Signals using Wasserstein Generative Adversarial Networks View Paper
Vlad Petrutiu, Liana-Daniela Palcu, Camelia Lemnaru, Mihaela Dinsoreanu, Rodica Potolea, Raul C. Mureșan and Vlad Moca
Abstract: Collecting EEG signal data during a human visual recognition task is a costly and time-consuming process. However, training good classification models usually requires a large amount of quality data. We propose a data augmentation method based on Generative Adversarial Networks (GANs) to generate artificial EEG signals from existing data, in order to improve classification performance on data collected during a visual recognition task. We evaluate the quality of the artificially generated signal in terms of the accuracy of a Convolutional Neural Network-based classifier that uses both real and augmented data to classify the outcome of the cognitive task. The preliminary results suggest that the introduction of artificially generated signals have a positive effect on the performance of the classifier. Moreover, we provide a method to quantify the level of information which indicates that the generated signals indeed follow the properties of the real ones.

Distributed Computing and Networking: Blockchain Applications (Friday, September 4, 10:00 - 11:20)
Channel: Track 02

Chair: Mariana Mocanu
Co-Chair: Tudor Cioara

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

10:00 - 10:20
Transactional Data Analytics for Inferring Behavioural Traits in Ethereum Blockchain Network View Paper
M S Bhargavi, Sushmitha M Katti, M Shilpa, Vaishnavi P Kulkarni and Supraja Prasad
Abstract: Ethereum is a blockchain based development platform for users to build and deploy decentralized applications and smart contracts. Transactions substituting money over Ethereum space is carried out using Ethereum’s cryptocurrency ‘Ether’. Though the decentralization and public ledger recording of the transactions proves its limpidity, the anonymity of the users, hiding their true identity behind the addresses echoes the need for discerning the behavioural traits in transactions. The appalling nature of Ethereum transactions to have both, security and threat at a comparable level, demands analytics for inferring traits for better perspectives and insights. This research work focuses on characteristic analysis of Ethereum transaction space for inferring behavioural traits in supervised and unsupervised context. In an unsupervised environment, raw transaction data is extradited to a tabular transaction structure by selecting appropriate features. The data is further clustered and validated to form coherent groupings of similar patterns. The clusters are characteristically analyzed based on the features through Radar plots for inferring behavioural traits. In supervised context, labelled transactions are represented using histograms and feature-based characteristic analysis is performed to infer traits. Such analytics lay foundations for future deeper analysis that can be used to discover trends and patterns to understand the transaction network, enhancement of trading strategies, identification of bot activities, anomaly detection and several others.

10:20 - 10:40
An Ethereum-based implementation of English, Dutch and First-price sealed-bid auctions View Paper
Claudia Pop, Mirela Prata, Marcel Antal, Cioara Tudor, Ionut Anghel and Ioan Salomie
Abstract: This paper addresses the problem of the traditional online auctions by proposing an Ethereum based implementation for three of the most popular auctions: English, Dutch, First-price sealed-bid (FPSB). Such centralized systems are subjects to malicious attacks, unreliable third-party for payment module, or repudiation of bids. We present a set of properties that advantages a blockchain auction, which tackles the drawbacks of a traditional one. Furthermore, we define smart contracts for expressing the auctions based on strict rules regarding the role of the participants and stage delimitation. While English and Dutch auction have a public character, the sealed bids from a FPSB can be questioned. The proposed solution incorporates the Merkle proof algorithm for honest verification of these offers. The experimental results show that a Dutch auction is more suitable for sellers, while an English auction for bidders in terms of transaction fees. Based on the behavior of each type of auction, the solution is exemplified by an English auction in the real estate field, including a protocol for storing data off-chain.

10:40 - 11:00
A Hyperledger integration for audit-enhanced decision support in a smart water distribution system View Paper
Diana-Andreea Arsene, Bogdan-Ionut Pahontu, Alexandru Predescu, Mariana Mocanu and Ciprian Lupu
Abstract: Water is one of the most important resources nowadays, which is why a good management can bring multiple benefits in modern society. With the continuous increase of water consumption, there is significant research regarding the efficiency of water control and monitoring process, in order to make optimal decisions for the proper functioning of a healthy eco-system. In the current paper we present a system that uses a water distribution network to evaluate the integration of modern concepts in a smart, secure and distributed approach for consumer demand monitoring. An IoT architecture is used for collecting data from smart meters, with push-based data exchange and reactive programming patterns. The network is integrated with a blockchain solution developed using Hyperledger Fabric to ensure an additional layer of security and transparency, while being used as an audit tool as well as data provider for decision support systems.

11:00 - 11:20
A Blockchain based solution for Managing Transplant Waiting Lists and Medical Records View Paper
Ioana Andreea Daniel, Claudia Pop, Ionut Anghel, Marcel Antal and Cioara Tudor
Abstract: Nowadays, the medical domain is suffering from a lack of transparency, security, and immutability. Developing a decentralized application based on blockchain technology can be helpful in achieving the requirements that will solve the identified problems. This paper proposes a blockchain-based solution for managing the waiting list for a transplant as well as a solution for sharing medical records between parties. Sharing medical records between patients and doctors represents a sensitive process since it must be secured, and it addresses a challenging requirement in terms of storage scalability which is approached by using a distributed file system. Results show that our solution allows the waiting list to be more trustful and increases the patients’ chances for receiving the transplant and at the same time assures the owner of the medical records, reducing the chances of attacks for stealing medical data.

Computer Vision: Automated Driving 2 (Friday, September 4, 11:30 - 12:50)
Channel: Track 03

Chair: Lucas Eiermann
Co-Chair: Florin Oniga

Mercedes-Benz AG, Germany
Technical University of Cluj-Napoca, Romania

11:30 - 11:50
Semantic 3D Obstacle Detection Using an Enhanced Probabilistic Voxel Octree Representation View Paper
Sorana Baciu, Florin Oniga and Sergiu Nedevschi
Abstract: In this paper we present a solution to overcome specific drawbacks of 3D obstacles detection and image based semantic segmentation, by combining the two approaches into an enhanced 3D octree. The 3D obstacle detection will be performed on the enhanced octree. Semantic information is added to the voxel octree representation in a probabilistic fashion, during the building step of the octree. Using the voxels' semantic labels and probabilities, more accurate obstacle primitives are extracted during octree traversal. Later, the semantic information helps for a better strategy in merging the 3D primitives. A better detection bounding box and pixel wise accuracy is achieved; as well a higher precision rate, as it allows easy introduction of 3D properties filtering. The evaluation metric is applied on KITTI's 3D object detection dataset. Semantic values are obtained using an already trained R-CNN model, while the 3D point cloud is generated with the semi-global block matching algorithm.

11:50 - 12:10
Trajectory Based Motion Planning for On-Ramp Merging-Situations Considering Extended Evaluation Criteria View Paper
Lucas Eiermann, Nick Bührer, Gabi Breuel, Ilja Radusch and Manfred Hauswirth
Abstract: In recent years, great progress in the area of automated and autonomous driving has been made. Nevertheless, key areas such as situation awareness and maneuvering still remain challenging, especially in urban areas and at highway junctions. In the following work a novel trajectory planner for automated lane merge situations will be presented. While our approach is adaptable to any merge scenario, we focus in this work on the most challenging maneuver, the on-ramp freeway merging. We present novel metrics for safety and comfort, such as merge-duration and lane usage and combine them with established key metrics such as timeto- collision, acceleration etc. Using these metrics, we validate our approach in a 3D physics simulation for two scenarios: dense highway traffic and highD-scenario based on real vehicle data.

12:10 - 12:30
Feasibility Study for an Automatic Architecture for Pothole Detection in Asphalt Images: a Trade-off between Performance and Quality View Paper
Leonardo Schmidt, Marcia Pasin and Sarajane Peres
Abstract: This paper presents a computational architecture to detect potholes in the asphalt images, classifying images into two sets: those that contain potholes and those that do not contain potholes. Two different classifiers -- Artificial Neural Network (ANN) and Support Vector Machine (SVM) -- were evaluated and their results compared against predefined metrics -- accuracy, precision, recall and F-score -- and also training and processing times. In our approach, in contrast to other works, we explore how parameter setting in classifiers influences the performance and the outcome quality. This feature is very important to assess the applicability of the solution in a real time system. Overall, we observed that SVM classifier allows better results than ANN, providing adequate rates across all metrics and evaluated configurations. On the other hand, ANN classifier requires more time for training, although the cost of training is not a concern in the recognition architecture with regard to the feasibility of real time application.

12:30 - 12:50
Autonomous Driving Architectures, Perception and Data Fusion: A Review View Paper
Gustavo Velasco-Hernandez, De Jong Yeong, John Barry and Joseph Walsh
Abstract: Over the last 10 years, huge advances have been done in the areas of sensor technologies and processing platforms, pushing forward developments in the field of autonomous vehicles, mostly represented by self-driving cars. However, the complexity of these systems has been also increased in terms of both hardware and software, especially for the perception stage which goal is to create reliable representation of the vehicle and the world. In order to manage this complexity, several architectural models has been proposed as guidelines to design, develop, operate and release self-driving solutions for real applications. In this work, a review on autonomous driving architectures is presented, classifying them into technical or functional architectures depending on the scope of their definition. In addition, the perception stage of self-driving solutions is analysed as a component of the architectures, detailing into the sensing part and how data fusion is used to perform localisation, mapping and object detection. Finally, the paper is concluded with additional thoughts on actual status and future trends in the field.

Intelligent Systems: Intelligence in Software projects (Friday, September 4, 11:30 - 12:50)
Channel: Track 01

Chair: Ciprian Dobre

Co-Chair: Radu Razvan Slavescu

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

11:30 - 11:50
The Impact of Punishment Postponement on the Evolution of Honest Behavior View Paper
Emil Bucur
Abstract: Interactions between honest and dishonest people are inevitable in society. Increasing crime rates raise concern among people and society leaders. A pressing and aware issue for more and more people is the importance of applying sanctions for criminal offenses on time. Postponing or delaying the punishment entails not only the increase of the criminal level and the ineffectiveness of the measures taken by the leaders of the society, but also a decrease in the level of people's trust in the act of justice. In the present paper we propose to study the impact of punishment postponement on the probability of punishment and how it can influence the level of honest/dishonest individuals in society. The purpose of this paper is to highlight the importance of timely sanctions for discouraging dishonest behavior. Our results may explain how punishment delay influences the mechanisms based on individual punishment.

11:50 - 12:10
Enhancing Scientific Collaborations using Community Detection and Document Clustering View Paper
Iulia-Maria Rădulescu, Ciprian-Octavian Truică, Elena-Simona Apostol and Ciprian Dobre
Abstract: Community detection is the process of extracting community structured subgraphs from community networks. Most research regarding community detection has focused on the network structure without taking the content associated with the nodes into account. In this paper, we propose a new method for enhancing a co-authorship network's structure using clustering. Specifically, considering the clustering process, we use a sequence with proved performance between the WordNet lemmatizer, Document Embeddings and Spherical K-Means, while choosing the Louvain algorithm for community detection. Thus, we improve the Louvain's community detection algorithm modularity by interconnecting the author nodes for the articles clustered together. To evaluate our method, we collected a dataset containing articles' abstracts and authors. The experimental results show that our method suggests potential new collaborations by adding vertices to the graph after analysing the textual content.

12:10 - 12:30
Interactive call graph generation for software projects View Paper
Istvan Attila Csaszar and Radu Razvan Slavescu
Abstract: This paper presents a tool aiming to help software developers visualize the call graph of their code while editing it. The tool can generate both the import graph and the call graph. It allows the user to select the nodes and edges to be displayed and, when s/he saves the document, it automatically updates the graph to reflect the current state of the project. The process of building the call graph or the import graph from the source code is based on using a query based architecture, commonly used by language servers. Performance benchmarks show that the tool can react to changes in the order of hundreds of milliseconds. To evaluate the tool usability, this was sent to 20 random users of which 9 decided to assess it. All respondents had an overall positive impression (gave a grade of 4 or 5 on a scale from 1 to 5) and also found the tool fast and easy to use.

12:30 - 12:50
Predicting Structure & Clarity of software projects with Machine Learning View Paper
Darius Mihnea Bogdan and Anca Marginean
Abstract: The software development domain is very dynamic and only around 28% of the projects that follow project management methodologies stay on schedule. This paper shows methods that try to solve this by aiming identification of the problems encountered by team members faster. The goal of the project is to allow project managers to find meaningful insights on the clarity and structure felt by team members on the tasks they are working on. The system we developed collects data from third party tools used for the development of software projects and correlates it with the marks given by team members for Structure & Clarity daily. We built models for prediction of Structure & Clarity using Random Forest Regression, Artificial Neural Networks (ANNs) and Long Short-Term Memory (LSTM).

Distributed Computing and Networking: Systems Quality and Safety (Friday, September 4, 11:30 - 12:50)
Channel: Track 02

Chair: Bela Genge

Co-Chair: Eneia Nicolae Todoran

George Emil Palade University of Medicine, Pharmacy,
Science and Technology of Targu Mures, Romania
Technical University of Cluj-Napoca, Romania

11:30 - 11:50
LOKI: A Lightweight Cryptographic Key Distribution Protocol for Controller Area Networks View Paper
Teri Lenard, Roland Bolboaca and Bela Genge
Abstract: The recent advancement in the automotive sector has led to a technological explosion. As a result, the modern car provides a wide range of features supported by state of the art hardware and software. Unfortunately, while this is the case of most major components, in the same vehicle we find dozens of sensors and sub-systems built over legacy hardware and software with limited computational capabilities. This paper presents LOKI, a lightweight cryptographic key distribution scheme applicable in the case of the classical in-vehicle communication systems. The LOKI protocol stands out compared to already proposed protocols in the literature due to its ability to use only a single broadcast message to initiate the generation of a new cryptographic key across a group of nodes. It's lightweight key derivation algorithm takes advantage of a reverse hash chain traversal algorithm to generate fresh session keys. Experimental results consisting of a laboratory-scale system based on Vector Informatik's CANoe simulation environment demonstrate the effectiveness of the developed methodology and its seamless impact manifested on the network.

11:50 - 12:10
Metric Semantics for Concurrent Languages Designed in Continuation-Passing Style View Paper
Eneia Nicolae Todoran
Abstract: We present a semantic domain for designing concurrent languages in continuation-passing style (CPS). We offer denotational semantics designed in CPS for languages based on Milner’s CCS extended with mechanisms which support multiparty interactions. We show that the basic laws of concurrent systems are satisfied in denotational semantics designed in CPS. The semantic investigation is accomplished in the mathematical framework of complete metric spaces.

12:10 - 12:30
The Quality of Programs Conceived by Object Enhanced Time Petri Nets View Paper
Dahlia Al Janabi, Mihai Florian Enache and Tiberiu Stefan Letia
Abstract: Software developers use metrics to estimate the software quality before the development stage to impact early decision making at the requirement stage, development quantification, operation, and software maintenance. These metrics give knowledge of a certain attribute of the software system and thus helps to evaluate it objectively and give opportunities for software improvement and making plans for modifications. The application development is supposed to be performed using the OETPN based modeling method. The newly proposed metrics are: Structural Complexity (number of arcs and the number of modules), Hierarchical Complexity (if one OETPN module is created by another), Unstructuredness of OETPN Models (which is based on the number of modules, cycles and decision places), Branching Complexity (which is computed according to the reachability graph), Decision Vertex (which is computed based on the decision node of the execution graph), and the Net Complexity is the summation of all the previous complexities.

12:30 - 12:50
A Physioteraphy Coaching System based on Kinect Sensor View Paper
Timeea Sărătean, Marcel Antal, Claudia Pop, Cioara Tudor, Ionut Anghel and Ioan Salomie
Abstract: This paper addresses the issue of performing physiotherapy treatments remote, at home, by proposing a personal coaching system capable of surveilling the correctness of the exercises and the adherence to the rehabilitation program established by doctors. For identifying the performed exercises, we use a Microsoft Kinect Sensor on top of which we have developed an exercise-detection algorithm part of a recognition module and a user-friendly web interface. The results show that the proposed solution can be successfully used for performing remote physiotherapy exercises, guided by physiotherapists, which brings a major advantage in scenarios like the one brought by the SARS-CoV-2 (COVID19) context.

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 1 (Friday, September 4, 16:00 - 17:20)
Channel: Track 03

Chair: Thomas Gatt
Co-Chair: Tiberiu Marita

The Malta College of Arts, Science and Technology, Malta
Technical University of Cluj-Napoca, Romania

16:00 - 16:20
MONet - Multiple Output Network for Driver Assistance Systems Based on a Monocular Camera View Paper
Razvan Itu and Radu Danescu
Abstract: Deep learning based image processing has become popular and approaches using convolutional neural networks (CNNs) have been widely used in recent years. In this paper we propose a multiple output convolutional neural network for road traffic scene understanding using a monocular camera. The color images are fed into the artificial neural network that produces multiple outputs. Our model performs three tasks: semantic segmentation, object detection and vanishing point computation. The semantic segmentation produces relevant data regarding the traffic scene, the obstacle detection module provides individual obstacles, whereas the vanishing point module will provide information that can be used to perform extrinsic camera calibration. We propose a novel obstacle detection approach and extend already published work by having a vanishing point detection module. The multiple outputs are predicted in a single-step and the information can be used as an initialization step for a 3D tracking system. Our network can extract individual dynamic objects and their correlation to the 3D space can be computed using the extrinsic parameters generated from the vanishing point module.

16:20 - 16:40
Detecting Highway Code Violations using Artificial Intelligence View Paper
Owen Xuereb, Frankie Inguanez and Thomas Gatt
Abstract: In this paper, we propose the use of autonomous law enforcement systems to aid in detecting highway code violations. Currently, the only way to enforce the law is to have an on-site official. This method tends to be very costly and allows multiple violations to go unnoticed. In turn, this reduces the efficacy of highway code enforcement. Automated law enforcement should fill this gap by reducing unnoticed contraventions which in turn contributes to safer driving and possibly reduces accidents. This research proposes the use of a Convolutional Neural Network (CNN) for object detection using two custom-made datasets for vehicles and STOP signs. This combination is used to identify a STOP sign and detect any vehicles passing over it without stopping. Our solution managed to detect 100% of STOP sign violations and non-violations in a real-life environment. The model for detecting the cars object and the STOP sign achieved an overall accuracy of 87.8% and 91.4% respectively.

16:40 - 17:00
Autonomous motor vehicle categorisation using a convolutional neural network View Paper
Christian Paul Cassar, Thomas Gatt and Ivan Briffa
Abstract: We present an automatic approach in recognising the classiﬁcation type of motor vehicles. Over the years, governments and private organisations have implemented schemes in which motor vehicles are categorised according to their size, weight, scale, and axel. Such schemes are used for various purposes, mainly for the implementation of new regulations, generation of statistics as well as urban development. This investigation proposes an autonomous system that is able to detect motorcycles, light motor vehicles (LMV’s) and heavy motor vehicles (HMV’s) from static images. This was established by training a Convolutional Neural Network (CNN) on preprocessed subsets of an existing dataset using the transfer learning technique. Our model achieved an F1 Score of 94.27% and an Accuracy Rate of 89.28% and challenging scenarios were identiﬁed.

17:00 - 17:20
Detecting road potholes using computer vision techniques View Paper
Neil Camilleri and Thomas Gatt
Abstract: The number of road potholes is growing day by day due to the increase in vehicles. This is also increasing the number of vehicle accidents caused by lack of road maintenance. In this study, a road pothole detection system using computer vision techniques was conducted. Object detection algorithms such as the YOLO, and SSD were selected. The YOLOv3-SPP model obtained the best mAP of 68.83\%, while the YOLOv3-tiny inferred an image in just 0.01s. Testing was also done on an Android device and a Raspberry Pi in order to evaluate performance on embedded systems. The SSDLiteMobileNet v2 converted to a TensorFlow Lite model outperformed all the YOLOv3 models when time is compared. In this paper we show that the model with the best performance is YOLOv3 SPP, but if time is a priority, the SSD TFLite model is the right fit.

Intelligent Systems: Natural Language Processing (Friday, September 4, 16:00 - 17:20)
Channel: Track 01

Chair: Adina Magda Florea
Co-Chair: Camelia Lemnaru

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

16:00 - 16:20
Speech Emotion Recognition method using time-stretching in the Preprocessing Phase and Artificial Neural Network Classifiers View Paper
Valentin Catalin Govoreanu and Mihai Neghina
Abstract: Human emotions are playing a significant role in the understanding of human behaviour. There are multiple ways of recognizing human emotions, and one of them is through human speech. This paper aims to present an approach for designing a Speech Emotion Recognition (SER) system for an industrial training station. While assembling a product, the end user emotions can be monitored and used as a parameter for adapting the training station. The proposed method is using a phase vocoder for time-stretching and an Artificial Neural Network (ANN) for classification of five typical different emotions. As input for the ANN classifier, features like Mel Frequency Cepstral Coefficients (MFCCs), short-term energy, zero-crossing rate, pitch and the speech rate were extracted. The proposed method was evaluated on the Ryerson Audio-Visual Database of Emotion Speech and Song (RAVDESS) and shows promising results when compared to other methods such as zero-padding.

16:20 - 16:40
Sentiment Analysis using Topic-Document Embeddings View Paper
Mădălina Mitroi, Ciprian-Octavian Truică, Elena-Simona Apostol and Adina Magda Florea
Abstract: Sentiment analysis plays an important role in automatically finding the polarity and insights of users with regards to a specific subject, events, and entity. In this article, we propose a new topic-document embedding (TopicDoc2Vec) for detecting the polarity of a text. The TopicDoc2Vec is constructed as the concatenation between a document embedding (Doc2Vec) and a topic embedding (Topic2Vec). A Doc2Vec is computed using word embeddings weighted by TF-IDF, while a Topic2Vec is computed using the word embeddings for the relevant terms belonging to the topic. To determine the quality of classification using the new proposed topic-document embeddings, we apply two algorithms, i.e., Logistic Regression and Long Short-Term Memory, for the task of sentiment analysis. We build the TopicDoc2Vec using five word embedding models, i.e., Word2Vec with CBOW and Skip-Gram, FastText with CBOW and Skip-Gram, and GloVe. The classification task is evaluated using accuracy, precision, and recall. For our empirical validation, we use a large labeled corpus containing game reviews. The experimental results prove that the TopicDoc2Vec embeddings outperform the Doc2Vec embeddings for detecting the polarity of a document.

16:40 - 17:00
Concept Identification with Sequence-to-Sequence Models in Abstract Meaning Representation Parsing View Paper
Orsolya Bernadett Batiz, Robert Paul Helmer, Roxana Pop, Florin Cristian Macicasan and Camelia Lemnaru
Abstract: A variety of NLP tasks require semantic information from sentence input. In order to capture meaning through Abstract Meaning Representation parsing, concept identification is required. In this research, the task of concept identification is approached using sequence-to-sequence models. Building upon an encoder-decoder architecture with attention mechanism, concepts are separated into verbs and non-verbs. Different alignment strategies are applied in order to increase the amount of train and test data for the model, resulting in a merged data-set with twice as many available instances. Furthermore, the training data increase induces an improvement of 10% for the sequence-to-sequence model.

17:00 - 17:20
A Two Stage Approach for AMR Parsing Using the Concept Inference Order View Paper
Roxana Pop, Florin Macicasan and Camelia Lemnaru
Abstract: In this paper we introduce a simple, pipeline graphbased approach for AMR parsing. Transition-based parsers have been researched intensively for the task of AMR parsing since the introduction of AMR and its associated sembank in 2013, but more recently there’s been a shift towards graph based approaches, which seem to do better on the task. However, these graph-based approaches often employ large models, trained end- to-end. We intend to see the limitations of a small, pipeline architecture. We use an attention-based sequence-to-sequence model for predicting an ordered list of nodes, then predict the heads for each node in a subsequent step.

HiPerGRID: Modern Streaming and Infrastructure (Friday, September 4, 16:00 - 17:20)
Channel: Track 02

Chair: Nicolae Tapus
Co-Chair: Ionut Anghel

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

16:00 - 16:20
A New Prefetching Scheme Based on Remote Control Inputs for Random Channel Change in IPTV Multicast Service View Paper
Hyunchul Joo, Intai Kim, Kyungchul Kwak and Dai-Boong Lee
Abstract: We propose an effective prefetching scheme for random channel change in IPTV multicast service. The proposed scheme determines a prefetching channel based on numeric key inputs of the remote control. It can reduce the channel zapping time with a little sacrifice of additional network bandwidth by performing procedures causing delay factors in parallel. Finally, experimental results are provided to show superior performance of the proposed scheme.

16:20 - 16:40
FreeBSD as a compute node in OpenStack View Paper
Andreea Mitran, Maria-Elena Mihailescu, Darius Mihai, Sergiu Weisz, Mihai Carabas and Nicolae Tapus
Abstract: At the moment, the demand for Infrastructure as a Service (IaaS) is higher than ever, as many companies from industry and open-source communities are focused on improving virtualization solutions. The aim of such a system is to provide virtual instances in a timely manner, persisting in quality and swiftness, while remaining simple and accessible for users. This paper proposes a model for a heterogeneous environment that offers services of IaaS using OpenStack as a cloud computing platform with Linux as a controller node and FreeBSD as a compute node. bhyve, the FreeBSD’s hypervisor, will be used to create the virtual instances.

16:40 - 17:00
RenderLink remote rendering platform for computer games View Paper
Bogdan-Ioan Oros and Victor Ioan Bâcu
Abstract: This paper aims to provide low latency video streaming for remotely rendered games based on a peer-to-peer model. Video streams are provided by peers that expose their hardware resources for rendering usage through RenderLink and are consumed by the web-native application provided by the backend service. The platform offers seamless integration of existing games (developed using GLFW3) with the platform. RenderLink approaches the subject of the high availability and scalability of the backend services used for the initial peer matching and connection establishment, as well as for a web-native interface for consuming game streams, through the usage of Kubernetes.

17:00 - 17:20
Particle Swarm Optimization Based Deep Learning Ensemble for Manufacturing Processes View Paper
Dorin Moldovan, Ionut Anghel, Tudor Cioara and Ioan Salomie
Abstract: Faults detection in semiconductor manufacturing processes is a challenging research problem because the manufacturing processes are characterized by hundreds of different types of operations and the identification of those operations that lead to faults in the final products is very challenging. The existence of faults during the manufacturing processes is strongly correlated with the performance of the manufacturing processes and in this article the main challenge that is approached is the accurate classification of the cases in which the manufacturing processes pass or fail the in line testing considering the information collected during the execution of the operations that are involved in the manufacturing processes. This article approaches the classification of the in line testing of the manufacturing processes using a method that is based on an ensemble of deep learning models in which the weights of the models are determined using Particle Swarm Optimization (PSO). The method is tested and validated on the open source SECOM dataset from the UCI Machine Learning Repository.

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 2 (Friday, September 4, 17:30 - 18:50)
Channel: Track 03

Chair: Pasi Pyykönen
Co-Chair: Mihai Negru

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

17:30 - 17:50
A Critical Evaluation of Aerial Datasets for Semantic Segmentation View Paper
Bianca-Cerasela-Zelia Blaga and Sergiu Nedevschi
Abstract: Drone perception systems use information from sensor fusion to perform tasks like object detection and tracking, visual localization and mapping, trajectory planning, and autonomous navigation. Applying these functions in real environments is a complex problem due to three-dimensional structures like trees, buildings, or bridges since the sensors (usually cameras) have limited viewpoints. We are interested in creating an application that is aimed towards inspection of forests with a focus on deforestation, with the main objectives being building a 3D semantic map of the environment and visual inspection of trees. In this paper, we evaluate three new datasets recorded at various flight altitudes, in terms of class balance, training performance on the semantic segmentation task, and the ability to transfer knowledge from one set to another. Our findings showcase the strengths of these datasets, while also pointing out their shortcomings, and offering future development ideas and raising research questions.

17:50 - 18:10
Narrowing the semantic gap between real and synthetic data View Paper
Radu Beche and Sergiu Nedevschi
Abstract: Image-to-image translation is an emerging method of computer vision dataset augmentation, which allows transferring the style of real life images onto synthetic ones, making them more realistic. In our work we propose an incremental improvement over the adversarial learning generator architectures used by image-to-image translation models. First, we managed to use a single network, instead of 2, thus creating a more memoryefficient model, which allowed for an end-to-end training on high resolutions. Second, inspired from recent work on semantic segmentation architectures, we enhanced our model by implying a multi-scale encoding and stylization phase, allowing for a better control over the contextual and spatial features. Given a synthetic image, our framework allows for its multimodal translation into the real domain. Our model shows promising results at narrowing the semantic gap between synthetic and real data.

18:10 - 18:30
Video Semantic Segmentation leveraging Dense Optical Flow View Paper
Vasile Lup and Sergiu Nedevschi
Abstract: Semantic segmentation is a key step in scene understanding for any autonomous system, as it provides the class of each pixel of an image. The best results on this task are achieved by deep neural networks, however they require in training a significant number of annotated frames which are difficult to create. This work is focused on improving the semantic segmentation network on sparsely labeled video sequences by including optical flow information between consecutive frames, and to provide a meaningful evaluation using densely labeled sequences. The system is composed of a static semantic segmentation, an optical flow and a linking network, which are chosen from existing architectures based on their high accuracy and low computational needs, but were never used together before. After conducting various experiments on the novel Virtual KITTI 2 dataset, we find that the optical flow compensates for the sparsity of annotations in video training sets, and by using just 25% of the labels we achieve the same video semantic segmentation quality as the static network trained on the fully labeled dataset. Since fewer annotations are needed, the cost of ground truth generation is reduced significantly. Moreover, taking advantage of the optical flow ground truth availability, we find the limits of the flow approach despite future advances, when more accurate optical flow networks will be developed. The system’s improved capabilities are also validated on the Cityscapes dataset.

18:30 - 18:50
Efficient spatio-temporal point convolution View Paper
Bogdan Maxim and Sergiu Nedevschi
Abstract: Real-time online semantic segmentation poses special challenges such as how to perform 3D convolution directly over the 3D geometric data which arrives in real-time. While different approaches exist, we propose a novel 3D convolution which operates directly on the point cloud, relies solely on the most important features that are found by the system. This eliminates the randomness proposed by other solutions, enabling only important information to be processed at runtime. Our architecture is powered by a novel heuristic search done on a two-level tree: a global interval tree and a local octree. While the global tree has the purpose of maintaining the 2D-3D information fusion, the local tree processes the neighborhood of local points into an ordered set, enabling execution of different operations. The trees update dynamically as new data arrives, while our heuristic search enables a better execution time, speeding the process from an average of 7 FPS to 11 FPS while maintaining accuracy. Through evaluation on public datasets, we show that our method gives results comparable to the state of the art, while gaining significant speed in terms of execution time.

Intelligent Systems: System Optimization (Friday, September 4, 17:30 - 18:50)
Channel: Track 01

Chair: Gheorghe Sebestyen
Co-Chair: Anca Marginean

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

17:30 - 17:50
Automated Machine Learning using Evolutionary Algorithms View Paper
Mihai Anton
Abstract: As the global data quantity already follows an exponential trend, machine learning has become present in every application, creating a great demand for general know-how, be it data scientists or computer scientists with related knowledge. Currently, the demand for work to be done surpasses the offer of such professionals, thus automatic solutions have to be found. The classical machine learning process involves data engineering, model selection, and hyperparameter tuning for the chosen model. Due to the highly repetitive nature of trial and error of these tasks, automation can play a big role in optimizing time spent on them. Automated Machine Learning comes to help the process by adding different optimization techniques that help data scientists be more productive and achieve similar or better results in a shorter time. This paper provides a novel approach to Automated Machine Learning using Evolutionary Algorithms and proves its performance by presenting top results in benchmark tests.

17:50 - 18:10
COVID-19 Preliminary Patient Filtering based on Regular Blood Tests using Auto-Adaptive Artificial Intelligence Platform View Paper
Zoltan Czako, Gheorghe Sebestyen and Anca Hangan
Abstract: In order to control the spread of the COVID-19 it is very important to identify those who have been already infected by this new type of virus. The rRT-PCR (reverse transcription polymerase chain reaction) testing is the golden standard for COVID-19 detection, but it is time consuming, laborious manual process and it is very short in supply. In order to reduce the number of tests, in this article we will present a possible solution for COVID-19 preliminary patient filtering based on regular blood tests, using artificial intelligence (AI) models. The most appropriate AI model will be selected using our auto-adaptive AI platform, AutomaticAI. The hyperparameters of the selected algorithm will also be adjusted automatically by this platform to match the context of the problem.

18:10 - 18:30
COVID-19 spread forecast using recurrent auto-encoders View Paper
Radu Beche, Romina Baila and Anca Marginean
Abstract: Since December 2019, the COVID-19 disease has become one of the most concerning issues across the Globe. The world has experienced serious prevention measures, from lockdown of cities to entire countries, as it is still unknown when will this pandemic end. Thus, it is no surprise that predicting the spread of this novel coronavirus has quickly become a hot topic among Artificial Intelligence researchers. In this paper we try to solve this task, by leveraging the capabilities that Recurrent Auto-Encoders have on time-series and implying a semi-supervised training process. Furthermore, the concept of nearest neighbour countries is introduced to estimate the cumulative number of confirmed cases for any country. The results are promising, showing that our proposed method is capable of making reliable predictions for a 30-days period. It is worth mentioning that, while this study uses information related to COVID-19, the proposed method can be used to estimate the evolution of any kind of disease, provided that the associated data comes in form of time-series.

18:30 - 18:50
Optimising the pool test method for COVID-19 using evolutionary algorithms View Paper
Cristian Lungu and Rodica Potolea
Abstract: The spread of the COVID-19 pandemic, quickly became a public health crisis which acted on many levels. The most challenging one of these was the sudden unavailability of protective gear and a complete lack of testing capacity. Although availability of masks and protective equipment has improved in the last few months, the testing capacity still remains a limited resource for most countries. One mitigation strategy for addressing the scarcity of tests is to pool biological samples in a single test, as demonstrated by the Frankfurt Goethe University. In this paper we add to the body of knowledge on the problem of optimizing the pooled testing strategy by optimizing a multistage adaptive testing scenario using an evolutionary algorithm. We also propose a generic framework by which optimizations can be advanced even further and will help to massively increase the testing capacity for stopping the current pandemic.

HiPerGRID: Satellite Data Processing (Friday, September 4, 17:30 - 18:50)
Channel: Track 02

Chair: Florin Pop

Co-Chair: Dorian Gorgan

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

17:30 - 17:50
Landscape Change Monitoring Using Satellite Data and Open Data Cube Platform View Paper
Catalin-Mihai Vajaiala-Tomici, Ion Dorinel Filip and Florin Pop
Abstract: This project aims to bring to the Romanian stage a system capable of providing information on various land changes that occur in Romania. These changes can have multiple causes, either natural or caused by human activities. Local authorities, business people, and researchers can use this information to make informed decisions that can influence Romanian citizens’ well-being. This paper is a starting point for building a complex open-source system at the national level, aiming to highlight the usefulness of using geospatial data. In order to interact in a pleasant way with geospatial data providers, this work is based on an open-source solution called Open Data Cube, which incorporates in a library different services, acting as an interface between the system itself and the geospatial data providers, but also as a standard for data storage. This paper explores three different use cases, namely the detection of deforestation, monitoring the evolution of urbanization, and the evolution of the Black Sea coast, especially in the area of the mouth of the Danube river. These three examples, although different, with totally different causes, highlight the versatility of such a system.

17:50 - 18:10
Exploration of a Visual Technique in Asteroids Detection Process View Paper
Denisa Copandean and Dorian Gorgan
Abstract: Starting from the Kuiper belt in the outer Solar System, covering the Main belt between Mars and Jupiter, ending up in the inner Solar System, the outer space surveys have gathered useful information about celestial objects which stand at the origin of life on Earth, asteroids. Yet, analyzing our planet, Earth, researchers have discovered that asteroids can also harm the life on Earth. Considering this fact, the interest in discovering and monitoring the Near Earth Asteroids (NEAs) has significantly increased in the last couple of decades. In the same time there was also a fast growth in technology allowing well-known surveys and amateurs mini-surveys to improve their solutions in order to obtain better results. In light of this situation, the automated pipeline for asteroids detection developed under the NEARBY project was designed in the form of a modular system. In this manner different asteroids detection techniques can be explored. Throughout this paper, a visual technique based on image processing will be investigated under different shades, as an alternative to the currently used astronomical technique, based on computation in the celestial reference system.

18:10 - 18:30
Qualitative Classification of Local Satellite Data View Paper
Razvan Margin and Dorian Gorgan
Abstract: The inference quality of machine learning solutions based on neural networks is highly related with the quality and quantity of training data. This paper proposes and experiments a set of algorithms and techniques to enhance the quality and quantity of training data based on satellite images. The results show that the algorithms can eliminate noise from different areas of interest, providing viable data for further processing. The achievements are assessed by the soil classification techniques on local Transylvanian areas.

18:30 - 18:50
A Multi-threaded Particle-in-cell Approach for Kinetic Plasma Simulations View Paper
Marius Joldos, Gabriel Voitcu, Alin Suciu, Anca Hangan, Marius Echim and Anca Marginean
Abstract: Particle-in-cell (PIC) simulations for plasma physics model self-consistently plasma phenomena at kinetic scales. Such simulations focus on the individual trajectories f a very large number of particles in self-consistent and external electric and magnetic fields, and are of great importance in astrophysics and space sciences. We developed a multi-threaded application to simulate such phenomena. This article describes the parallel performance of our application on a high-performance multicore system

18:50 - 19:10
Development and Research on CLOUDUT Infrastructure
Dorian Gorgan and Sergiu Nedevschi
Abstract: CloudUT

Intelligent Systems: Knowledge and Rule Based Systems (Saturday, September 5, 09:00 - 10:20)
Channel: Track 01

Chair: Viorel Negru
Co-Chair: Adrian Groza

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

09:00 - 09:20
A decision support system for B2B CRM Systems using belief rules View Paper
Doru Rotovei, Daniela Zaharie, Viorel Negru and Daniel Pop
Abstract: The complexity of business environment often forces sales representative to take decisions using intuition and/or subjective models created through experience instead of data driven decisions. However, data driven decisions generate better revenue. In this work, we present a methodology to create a Decision Support System (DSS) for Business to Business (B2B) sales that generates recommendations for actions to follow and actions to avoid to increase the chances of successfully closing a deal. At the core of our proposed DSS, is a three-step methodology relying on a particular approach of using belief rules. In the first step we generate, based on available data on previous won/lost deals, some forecasting models. As the best performing models are of black box type, the second step aims to extract if-then conventional rules that are converted, in the last step, into belief rules. Rather than using an attribute value probability matching to activate a belief rule, we propose a cost of activation that takes into account the current state of the active deal and the cost of transitioning the active deal's attributes into a new state that could activate a belief rule. Furthermore, we propose and analyze two methods of calculating the cost of transition related to the effort the sales representative should dispose to move the current deal into the new state. The three-step approach was used to create a DSS for a real world B2B complex sales Customer Relationship Management Systems and the results are discussed and included in this paper.

09:20 - 09:40
onIQ: An Ontology-Independent Natural Language Interface for Building SPARQL Queries View Paper
Ilie Cristian Dorobat and Vlad Posea
Abstract: Publishing data in Linked Data tends to capture more and more space because this way of representation of data allows linking the data, which leads to an increase in the degree of representation and, why not, the quality of data. Unfortunately, Linked Data has a much more complex and heterogeneous structure than other ways of data representation, which means only a larger learning curve of SPARQL, the triplestore query language. In this paper, we propose a framework, Ontology-Independent Natural Language Query Interface for Building SPARQL Queries, for translating queries formulated in natural language into SPARQL queries, so as to facilitate the adoption of Linked Data. The framework uses spaCy library for applying specific NLP techniques for syntactic analysis and extraction of the elements needed to build the SPARQL query. Through WordNet, the synonyms for the words extracted from the query user are identified, and then, in order to map them with properties from the used ontologies, the degree of similarity is calculated. Finally, the SPARQL query is built and the user can use it for querying the triplestore.

09:40 - 10:00
Fake news detector in the medical domain by reasoning with description logics View Paper
Adrian Groza and Ana Diana Pop
Abstract: We investigate here how reasoning in Description Logics (DLs) can detect inconsistencies between trusted medical sources and not trusted ones. The not-trusted information comes in natural language (e.g. ”Depression is not an illness”). To automatically convert into DLs, we used the FRED converter. Reasoning in Description Logics is then performed with the Racer and Hermit reasoners to detect inconsistencies. Each inconsistency corresponds to a possible medical myth. The detected inconsistencies are verbalised to the human agent aiming to explain why a sentence is considered a medical myth. Our fake news detector is exemplified on in the psychocognitive and epidemiological fields.

10:00 - 10:20
MineFOL: a Game for Learning First Order Logic View Paper
Adrian Groza, Maria Monalisa Baltatescu and Mihai Pomarlan
Abstract: The role of games in computing science has been given some consideration in the literature. We aim to design games that can be used for teaching logic. Specifically, we build on the Minefield game used in the Stanford Intro to Logic course, and present a version, which we call MineFOL, that incorporates time constraints. We also move beyond the Stanford examples in that we provide a formalisation of the game, with a view to eventually generating game instances automatically as well as from user input, and show how \m can be used as a learning tool. We also developed an online platform supporting learners in at least three ways. First, the learners can practice reasoning in First Order Logic and proving strategies such as resolution through various \m games. Second, the learners can practice formalisation in FOL, by allowing learners to build their own games. Third, the learners become aware of the need of interleaving various technologies to solve a logic-based task. In this line, \m is formalised in such way to allow augmenting reasoning in FOL with searching strategies.

Intelligent Systems: Applications (Saturday, September 5, 09:00 - 10:20)
Channel: Track 02

Chair: Ioan Salomie
Co-Chair: Cristian Vicas

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

09:00 - 09:20
Evaluation of Deep Learning architectures for System Identification. Applications for an household heating system View Paper
Cristian Vicas
Abstract: Model Predictive Controllers are widely used in the industry to control a dynamic system. They require a good model for the controlled process. In this paper I try to perform system identification using promising current neural network architectures. I frame the problem as a regression setup but allow for future possible commands to be integrated in the response. I test six architectures on three datasets. I adapt successful neural network architectures from Computer Vision and Natural Language Processing to predict the output of a system for some future time steps, given current knowledge plus some future commands. Results are promising but the models are not ready to be integrated in a real time, resource constrained environment.

09:20 - 09:40
Human Activity Recognition using Smartphone Sensors and Beacon-based Indoor Localization for Ambient Assisted Living Systems View Paper
Andreea Valeria Vesa, Vlad Simion, Renato Rus, Marcel Antal, Claudia Pop, Ionut Anghel, Tudor Cioara and Ioan Salomie
Abstract: The aim of this paper is to present a novel indoor human activity recognition system designated for ambient assisted living environments. To overcome the drawbacks of current solutions in terms of equipment costs and computational complexity, we propose a simple yet robust activity recognition system that combines the output of an activity classification algorithm using smartphone sensors such as accelerometer and gyroscope with a Beacon-based indoor localization predictor to detect more complex activities based on a decision rule system. For the multiclass classification algorithm we employ a ConvLSTM algorithm, while for the positioning system we propose an ensemble based solution combining Multilayer Perceptron with Gradient Boosted Regression and k Nearest Neighbors. Our solution has been tested in a controlled home environment setup achieving an average localization error of 0.4m while for the final activity recognition system we report an accuracy of 91.0% concluding that the proposed solution is accurate enough to be integrated as a monitoring tool facility within ambient assisted living systems.

09:40 - 10:00
A System for Managing and Processing Industrial Sensor Data: SMS View Paper
Chira Codrin-Mihai, Raluca Portase, Ramona Tolas, Camelia Lemnaru and Rodica Potolea
Abstract: The business value embedded in the IoT devices data is estimated as precious as the difficulty to seize it. Even though the costs of accumulating and analysing this information frequently outweigh the returned avail, the scope of the described solution is to overcome these complexities. The modernisation that arose around the Industry 4.0 context has generated the demand for a competing system competent in streamlining the methods and mechanisms of big data administration. The system’s flexibility and adaptability are essential characteristics in such a progressive domain. The purpose of our work is to define and implement a full flow processing pipeline able to adapt to the outlined trends. The technology and design-related choices on which the architecture is developed are examined based on the theoretical grounds identified for the specific industrial challenges.

10:00 - 10:20
Modern approaches to preprocessing industrial data View Paper
Eliza Olariu, Ramona Tolas, Raluca Portase, Mihaela Dinsoreanu and Rodica Potolea
Abstract: In the context of the current technological progress, big data arises as a compelling research topic. This paper presents a taxonomy of non-traditional analysis strategies for industrial data within a Big Data paradigm. A general approach of preprocessing operations for exploring and exploiting valuable knowledge from large set of industrial data is defined. The identified strategies are tested and validated on a real industrial data set characterized by a multitude of complexities. The methodology involves syntactic and semantic processing steps including building various data visualization mechanisms and correlating the signals. Each of the presented approaches handles a real data complexity.

Computer Vision: Technologies and Applications - 1 (Saturday, September 5, 09:00 - 10:20)
Channel: Track 03

Chair: Stefan Holban
Co-Chair: Raluca Brehar

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

09:00 - 09:20
Improving the Accuracy of Deep Neural Networks Through Developing New Activation Functions View Paper
Marina Adriana Mercioni, Angel Marcel Tat and Stefan Holban
Abstract: Without activation functions, it would be only possible for the neural network to learn very basic tasks, so the activation function is a key point in the neural network’s architecture. The function allows us to learn more complicated tasks and also it impacts the performance to obtain the outcome. So, activation functions represent the continuous and widespread interest of research to identify the most suitable activation function to a specific task. In this paper, we propose four activation functions that bring improvements for different datasets in the Computer Vision task. These functions are a combination of the popular activation functions such as sigmoid, bipolar sigmoid, Rectified Linear Unit (ReLU), and tangent (tanh). By allowing activation functions to be learnable we obtain models more robust. To validate these functions, we tested using more datasets and more architectures with different depths, showing that their properties are significant and useful. Also, we compared them with other powerful activation functions to see how our proposed activation functions impact accuracy.

09:20 - 09:40
Shape and Contrast Adaptive Non-local Means Filter View Paper
Claudiu Bic, Romulus Terebes and Raul Malutan
Abstract: In this paper we propose an improvement to the NLM algorithm. It consists of an adaptive threshold parameter used to compute the similarity function of the Non-Local Means (NLM) approach integrated in the NLM-SAP (shape-adaptive-patches) algorithm. To obtain the adaptive threshold parameter we use an algorithm that decomposes the image in a textural and a homogenous part. Additionally, a prefiltering method is used on the input noisy image, that is based on a discrete version for TV filtering. Experimental results show that this allows the obtention of better visual, and objective measures-quantified results than our previous work and other considered NLM approaches. In some cases, better visual results were obtained compared with the state-of-the-art BM3D (Block Matching 3D) algorithm and the results in terms of PSNR/MSSIM are comparable.

09:40 - 10:00
Face Retrieval in Videos using Face Quality Assessment and Convolution Neural Networks View Paper
Rahma Abed, Sahbi Bahroun and Ezzeddine Zagrouba
Abstract: With the large amount of videos produced every day, Content-Based Video Retrieval (CBVR) has become a necessity by describing each video with a compact and significant signature in order to efficiently retrieve the desired video from a large collection. In this work, we present a CBVR system applied on face recognition based on keyframes. The first step in this system consists to extract keyframes from video using Face Quality Assessment and Convolution Neural Networks. Starting by generating face quality scores for each face image using three face feature descriptors (Gabor,Local Binary Pattern (LBP) and Histogram of Oriented Gradient (HoG)). Then, we train a Convolution Neural Network (CNN) in a supervised manner in order to select frames having the best face quality. Experiments on several datasets has shown that the proposed “DeepFQA” method gives a promising results in terms of accuracy and precision/recall curve.

10:00 - 10:20
Driver Drowsiness Detection Based on Joint Monitoring of Yawning, Blinking and Nodding View Paper
Aicha Ghourabi, Haythem Ghazouani and Walid Barhoumi
Abstract: Deaths and injuries resulting from road traffic crashes remain a serious problem globally and current trends suggest that this will continue to be the case in the foreseeable future. In this paper, we propose a reliable method towards drowsiness detection by analyzing images of the driver’s face. In fact, the shadows caused by wearing glasses and/or bad light conditions may, in particular, decrease the accuracy rate of the blinking detection. In addition, drowsiness symptoms are not just restricted to the frequency of blinking, as most of existing works do, but also include yawning and nodding. Thus, instead of using a single facial feature to predict the drowsiness state, we jointly combine eye closure and yawning by measuring the eye and the mouth aspect ratios, and head pose which is estimated by analyzing the optical flow. Then, we investigate the multilayer perceptron and the K-nearest neighbors as two classification techniques. The proposed method has been tested on the challenging private NTHU-DDD benchmark video dataset, and the preliminary experimental results show the effectiveness of the proposed automated method.

Intelligent Systems: Learning (Saturday, September 5, 10:30 - 11:50)
Channel: Track 01

Chair: Emil Stefan Chifu
Co-Chair: Calin Cenan

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

10:30 - 10:50
Conformal transformation of the metric for k-nearest neighbors classification View Paper
Claudiu Popescu, Lacrimioara Grama and Corneliu Rusu
Abstract: The paper introduces a new method for improving the nearest neighbors classifier. Our approach is based on the idea of replacing the constant metric with a variable and conformally equivalent one that is data dependent, and therefore it is more informative. We define a family of conformal transformations that, under some assumptions, induces distance functions that are efficiently computable. Using the intuition that the distances between points near a class boundary should be larger, a simple method for selecting a transformation is proposed. We perform experiments on two datasets. The first set of experiments are with a sentiment prediction dataset, and in this case our method offers some improvements over the standard k-NN algorithm. In the second empirical analysis, we apply the method to a news categorisation problem. In this case the results are mixed. We conclude with a discussion of the advantages and weaknesses of the method, and propose a number of possible improvements.

10:50 - 11:10
A Graph-Based Model for Combinatorial Auctions View Paper
Paul Diac
Abstract: In this paper, we propose a model for bidding in combinatorial auctions and an allocation algorithm for the model. Both are integrated in an application enabling expressive and concise bids and computing the winner determination, i.e., optimal allocation. Combinatorial auctions are a special, more expressive format for organizing auctions based on complex user preferences, motivated by the limitations of the traditional, single-item valuations. In practice, often items would be valued in conjunction with others, in pairs, groups; based on their quantity, and other possible dependencies. Such valuations are declared in our application based on LISP functions, labeling edges of a special flow graph. The allocation algorithm scans these valuations to guess the best possible allocation with partial information.

11:10 - 11:30
Enhancing Tokenization by Embedding Romanian Language Specific Morphology View Paper
Mihaela Alexandra Vasiu and Rodica Potolea
Abstract: This paper addresses the significance, complexity and approaches of Romanian tokenization, the first step in NLP. A strategy is proposed for tokenizing the Romanian text with various specificity: standard (literal language), non-standard (archaic) and medical. The strategy is extending existing tokenization models for English, Spanish and Portuguese with Romanian features. Moreover, we attempt to address the Romanian ambiguities caused by the complexity of the grammar and provide support for further Romanian NLP processing. The approach proved that there is no general solution for tokenization and that diverging tokenization strategies for all languages can lead to considerably better results. With the current approach we obtained a high-quality tokenization on all three Romanian formats as follows: <97.01% precision, 96.84% recall> for standard sources, <98.10% precision, 97.03% recall > for non-standard sources and <94.73% precision, 95.36% recall > for medical sources.

11:30 - 11:50
Round Based Extension Algorithm for Gossip Learning View Paper
Szabolcs Bagoly and Radu Danescu
Abstract: Distributed training algorithms represent a promising direction in providing specially crafted machine learning models, personalized on more appropriate user data samples producing greater overall user experience. We propose an extension for the fully distributed, peer to peer based Gossip Learning algorithm, enhancing it with additional memory for storing local caches of model updates. In our simulations, the extension algorithm achieved better performance for both accuracy and convergence time. One inherent advantage of the extension is that it makes the algorithm better suited for deployment on mobile devices. We also examine some technical aspects of deploying a Gossip Learning framework onto mobile devices, defining some basic building blocks, followed by technical implementation solutions for the Android operating system.

Intelligent Systems: Security and Privacy (Saturday, September 5, 10:30 - 11:50)
Channel: Track 02

Chair: Robert Varga
Co-Chair: Ciprian Oprisa

Technical University of Cluj-Napoca, Romania
Bitdefender, Romania

10:30 - 10:50
Detecting Injection Attacks using Long Short Term Memory View Paper
Ioana-Crinela Potinteu and Robert Varga
Abstract: Nowadays, the rise of cybercrime requires the companies to empower their business with innovative defensive mechanisms. This article focuses on the security of RESTful APIs and it proposes two methods to defend them against different types of injection attacks. The analyzed APIs belong to a large scale company and they respond to more than 10 000 requests per second, coming from applications that belong to totally different business domains. At the foundation of these solutions is the usage of the powerful Long Short Term Memory network and two different tokenization strategies: a word-level and a character-level. The solutions outperform the current model used by the company and are able to detect an injection attack tentative sent by a client with 96% accuracy, in comparison with the model employed, which only achieves 92% accuracy.

10:50 - 11:10
Towards Pentesting Automation Using the Metasploit Framework View Paper
Ovidiu Valea and Ciprian Oprisa
Abstract: Penetration testing is a well known methodology assessing security vulnerabilities by executing complex steps which form an attack. Professional pentesting is an expensive service that sometimes cannot fit in the budget of Small and Medium Enterprises. Automating this process means it can be executed even by inexperienced system administrators while it saves time for professionals. The difficulty of this problem consists in the heterogeneity of networks and systems so the techniques need to be adapted each time. Our approach is based on identifying system characteristics, search for existing vulnerabilities and applying machine learning for selecting the most appropriate exploit. The model was trained using data collected from exploited machines on the "Hack the Box" learning platform and delivers exploits from the Metasploit framework. The evaluation shows that the proposed framework can exploit a fair number of systems and can be extended to support new classes of exploits and new pentesting methodologies.

11:10 - 11:30
Privacy Oriented Ecosystem Comparable to Google’s Ecosystem View Paper
Emanuel Truța and Alin Suciu
Abstract: As technology evolved, questions about privacy became more and more prominent "What is Google doing to my data?", "What do they know about me?", "How to keep safe?", are just a few questions that arise. Can we trust Google with our data? Ultimately, data that is not in our hands is not ours to manage either. They do whatever they want with it, and there's no stopping them, unless we don't give them our data in the first place. This paper aims to dive deep into how tech companies make use of your data, what they can potentially do with it, what is known they do with it, and to provide an alternative that compromises as little data as possible while keeping as much functionality as possible. As we will see, many times, there is a tradeoff between privacy and functionality.

11:30 - 11:50
Detecting Malicious URLs Based on Machine Learning Algorithms and Word Embeddings View Paper
Andrei Crisan, Gabriel Florea, Lorand Halasz, Camelia Lemnaru and Ciprian Oprisa
Abstract: Relying on the appropriate features is essential in classification models for malware detection, for various important reasons, such as dealing with class imbalance, the ability to detect zero-day malware samples, or preventing attackers to successfully reverse engineer the classification process and changing non- essential feature values to avoid detection. In this paper, we propose a method that uses a combination of word embeddings together with ”classical”, domain-engineered features, to obtain reliable classification models for malicious URLs detection. Ad- ditionally, we explore different traditional techniques to address class imbalance – such as synthetic oversampling or cost-sensitive learning – and several classification techniques. We find that the best overall results are obtained by using a cost-sensitive neural network – with a precision that exceeds 99% and an accuracy above 90%, while maintaining a recall rate above 89%. We have performed an analysis of the importance of the features proposed, and found that while word embeddings produce better results than bi-gram based features, domain-specific features are necessary for obtaining a high precision in detecting malicious URLs.

Computer Vision: Technologies and Applications - 2 (Saturday, September 5, 10:30 - 11:50)
Channel: Track 03

Chair: Joseph Walsh
Co-Chair: Ion Giosan

Institute of Technology Tralee, Ireland
Technical University of Cluj-Napoca, Romania

10:30 - 10:50
A Comparison of Deep Learning Models for Pallet Detection in Industrial Warehouses View Paper
Michela Zaccaria, Riccardo Monica and Jacopo Aleotti
Abstract: We consider the problem of automatic pallet detection in industrial environments using a single RGB camera. The problem is relevant in the context of autonomous guided vehicle navigation, and for tasks like pallet storage and retrieval. In particular, we present an approach based on a convolutional neural network (CNN) followed by a decision making step. The convolutional neural network is trained to recognize two elements of a pallet, namely the pallet front side and its pockets. We also report a comparison between three state-of-the-art CNNs: Faster R-CNN, SSD and YOLOv4. For training and evaluation, a dataset was collected in a warehouse. The dataset contains images of pallets in different configurations, either on the ground or on racks, and with arbitrary orientation.

10:50 - 11:10
Automatic Vision Inspection Solution for the Manufacturing Process of Automotive Components Through Plastic Injection Molding View Paper
Mircea Paul Muresan, Dragos Gabriel Cireap and Ion Giosan
Abstract: In the automotive industry, vehicle components can be obtained through the process of plastic injection molding. The components can be fixed in the vehicle by using metal bushings, which are placed on the injection mold before the beginning of the plastic injection process. The incorrect placement or absence of the bushings leads to a defective product. Object classification has been a long-tackled problem in Computer Science, and the breakthroughs in Artificial Intelligence allows us to solve this problem with a high degree of accuracy and precision. In the context of industrial inspection, ensuring high-quality products is a matter of utmost importance. The automated vision inspection process facilitates the creation of a high volume of products, which are in conformity with the quality standards, in a short amount of time and can save the manufacturing company money. In this paper we propose a solution that automatically detects the injection mold and classifies the positioning of the bushings, warning the operator in case the current positioning can lead to a defective product. Furthermore, the bushing detection is optimized in such a way that, it is mandatory for the mold to be visible only in the first frame, thus reducing the running time of the whole pipeline. The results are validated by using a data set covering multiple possible working scenarios. Moreover, we compare the implemented classifier with other classifiers, highlighting the performance of the proposed solution with respect to running time and classification accuracy.

11:10 - 11:30
Single-Ray LiDAR-based Scanner for Static Scenes Dense 3D Reconstruction View Paper
Bogdan Mesesan, Ramona Voichita Galatus and Tiberiu Marita
Abstract: This paper summarizes the design of the electro-mechanical and software implementation of a single-ray LiDAR- scanner system for the dense 3D reconstruction of static scenes. A test bench was designed aiming to acquire precise distance measurements, using a single-ray Garmin Lite V3 LiDAR, and transform them into a high-resolution point cloud. The goal is to demonstrate the algorithm behind the laser: time-of-flight based 3D reconstruction technique. The yawing scan and pitching scan methods (rotation of the sensor around two orthogonal axis (yaw and pitch)) are used for the current application. A microcontroller application based on the Arduino Uno board was designed and implemented to acquire data from the LiDAR at different positions. Since the LiDAR is a one-dimensional sensor, the electro-mechanical setup must be capable of achieving the second and third dimensions. The information from the scanner is collected based on the spherical coordinates: θ - polar rotation angle (yaw), φ – azimuthal angle (pitch) and r – Euclidian distance. An accelerometer based PID control algorithm was developed to set the sensor at an initial azimuthal angle, prior to the scan. A desktop application was designed to collect the information from the sensor. It includes a graphical user interface used for setting the scan parameters or commencing a scan routine. A data processing application takes the acquired measurements and calculates the Cartesian coordinates of each pixel based on the parameters of the scan. A 3D reconstruction of the scan field of view is generated. The systematic errors of the system have also been calculated using a static reference scenario. Additional experiments in several static scenarios were performed for the quantitative and qualitative assessment of the results.

11:30 - 11:50
Intersection Management Systems and Internet of Things: A Review View Paper
Gustavo Velasco-Hernandez, Eduardo Caicedo-Bravo, John Barry and Joseph Walsh
Abstract: Intersections or junctions are critical points in transportation systems due to their dynamic behaviour, making them prone to safety or efficiency problems. For this reason, the use of technology in monitoring intersections has increased over the last years. In addition, different proposals based on the Internet of Things (IoT) have been used to handle different traffic issues at intersections, taking advantage of the features it provides regarding hardware, software and communications. This work presents a comprehensive review on Intersection Management Systems (IMS), from a component-based point of view, and a special remark on how IoT-based solutions have been adopted recently in this area. A proposed architecture for the implementation of an IoT-ready IMS is presented as well.

Computer Vision: Medical Imaging - 3 (Saturday, September 5, 12:00 - 13:20)
Channel: Track 03

Chair: Radu Badea

Co-Chair: Delia Mitrea

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

12:00 - 12:20
Adversarial Graph Learning and Deep Learning Techniques for improving diagnosis within CT and Ultrasound images View Paper
Dan-Sebastian Bacea, Delia-Alexandrina Mitrea, Sergiu Nedevschi and Radu Badea
Abstract: Computed tomography (CT) examination has an important role in diagnosing, monitoring, and evaluating treatment response in patients infected with COVID-19. Chest CT images provide a reliable alternative to the reverse transcription polymerase chain reaction (RT-PCR). Ultrasonography provides a non-invasive and low-cost solution for diagnosing liver tumors. Recognition of hepatocellular carcinoma in ultrasound images can provide a suitable alternative to the more invasive and dangerous method of doing a biopsy. Significant effort has been dedicated to the development of automated ways of diagnosing COVID-19 based on CT scans and on diagnosing liver tumors based on ultrasound images. Using deep learning methods shows promising results. This paper proposes a method for training deep neural networks for performing binary classification, which is called Adversarial Graph Learning. Using Adversarial Graph Learning leads to improved results and to more robust learned features on both target tasks of recognizing COVID-19 in chest CT images and of recognizing hepatocellular carcinoma in ultrasound images.

12:20 - 12:40
Teeth Detection and Dental Problem Classification in Panoramic X-Ray Images using Deep Learning and Image Processing Techniques View Paper
Mircea Paul Muresan, Andrei Răzvan Barbura and Sergiu Nedevschi
Abstract: Deep convolutional neural networks, have gained a lot popularity in medical research due to their impressive results in detection, prediction and classification. Analysis of panoramic dental radiographies help specialists observe problems in poor visibility areas, inside the buccal cavity or in hard to reach areas. However, poor image quality or fatigue can cause the diagnosis to vary, which can ultimately hinder the treatment. In this paper we propose a novel approach of automatic teeth detection and dental problem classification using panoramic X-Ray images which can aid the medical staff in making decisions regarding the correct diagnosis. For this endeavor panoramic radiographies were collected from three dental clinics and annotated, highlighting 14 different dental issues that can appear. A CNN was trained using the annotated data for obtaining semantic segmentation information. Next, multiple image processing operations were performed for segmenting and refining the bounding boxes corresponding to the teeth detections. Finally, each tooth instance was labeled and the problem affecting it was identified using a histogram-based majority voting within the detected region of interest. The implemented solution was evaluated with respect to several metrics like intersection over union for the semantic segmentation and accuracy, precision, recall and F1-score for the generated bounding box detections. The results were compared qualitatively with the data obtained from other approaches illustrating the superiority of the proposed solution.

12:40 - 13:00
3D Texture Feature Extraction and Classification using the BM3DELBP approach View Paper
Stefania Barburiceanu, Romulus Terebes and Serban Meza
Abstract: This paper presents a new feature extraction technique for 3D texture classification based on Local Binary Patterns-derived operators and state-of-the-art denoising techniques, with good invariance properties to different transformations and robust to noise. Our approach consists in the computation of textural features obtained using the Block Matching and 3D Filtering Extended Local Binary Patterns for each of the three orthogonal planes and their concatenation. The proposed method achieves a promising classification performance on a publicly available dataset of 3D textured images. The obtained results prove that the presented feature extraction technique is above traditional 2D texture descriptors and typical CNN deep learning-based methods.

13:00 - 13:20
A Two Step Approach for Joint Detection View Paper
Irina Maria Sandu, Irina Mocanu and Lucia Rusu
Abstract: Human joints can be used in areas such as activity recognition, animation, augmented reality and robotics. Thus, there is a need to correctly detect these points. This paper proposes a two step method for joint detection. First an initial prediction is performed using the HRNet network. In the second step an improvement is applied on the initial predictions for each joint using a convolutional neural network. Tests were performed using the Unite the People dataset. The new model is able to improve the results of the base model by 1.81%.

Thursday, September 3, 2020

Friday, September 4, 2020

Saturday, September 5, 2020

Detailed Technical Program 2020

Workshops

Bosch Student Workshop (Thursday, September 3, 14:00 - 15:50) Channel: Track 03

Integrated Semantic Visual Perception and Control for Autonomous Systems (Thursday, September 3, 16:10 - 18:00) Channel: Track 03

FROM 2020 (Thursday, September 3) Channel: TBA

Keynote Lectures

Plenary Presentation 1 (Friday, September 4, 09:00 - 09:50) Channel: Track Plenary

Plenary Presentation 2 (Friday, September 4, 14:00 - 14:50) Channel: Track Plenary

Plenary Presentation 3 (Friday, September 4, 15:00 - 15:50) Channel: Track Plenary

Conference Sessions

Computer Vision: Automated Driving 1 (Friday, September 4, 10:00 - 11:20) Channel: Track 03

Intelligent Systems: Fuzzy Reasoning and Brain Signal Classification (Friday, September 4, 10:00 - 11:20) Channel: Track 01

Distributed Computing and Networking: Blockchain Applications (Friday, September 4, 10:00 - 11:20) Channel: Track 02

Computer Vision: Automated Driving 2 (Friday, September 4, 11:30 - 12:50) Channel: Track 03

Intelligent Systems: Intelligence in Software projects (Friday, September 4, 11:30 - 12:50) Channel: Track 01

Distributed Computing and Networking: Systems Quality and Safety (Friday, September 4, 11:30 - 12:50) Channel: Track 02

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 1 (Friday, September 4, 16:00 - 17:20) Channel: Track 03

Intelligent Systems: Natural Language Processing (Friday, September 4, 16:00 - 17:20) Channel: Track 01

HiPerGRID: Modern Streaming and Infrastructure (Friday, September 4, 16:00 - 17:20) Channel: Track 02

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 2 (Friday, September 4, 17:30 - 18:50) Channel: Track 03

Intelligent Systems: System Optimization (Friday, September 4, 17:30 - 18:50) Channel: Track 01

HiPerGRID: Satellite Data Processing (Friday, September 4, 17:30 - 18:50) Channel: Track 02

Intelligent Systems: Knowledge and Rule Based Systems (Saturday, September 5, 09:00 - 10:20) Channel: Track 01

Intelligent Systems: Applications (Saturday, September 5, 09:00 - 10:20) Channel: Track 02

Computer Vision: Technologies and Applications - 1 (Saturday, September 5, 09:00 - 10:20) Channel: Track 03

Intelligent Systems: Learning (Saturday, September 5, 10:30 - 11:50) Channel: Track 01

Intelligent Systems: Security and Privacy (Saturday, September 5, 10:30 - 11:50) Channel: Track 02

Computer Vision: Technologies and Applications - 2 (Saturday, September 5, 10:30 - 11:50) Channel: Track 03

Computer Vision: Medical Imaging - 3 (Saturday, September 5, 12:00 - 13:20) Channel: Track 03

Bosch Student Workshop (Thursday, September 3, 14:00 - 15:50)
Channel: Track 03

Integrated Semantic Visual Perception and Control for Autonomous Systems (Thursday, September 3, 16:10 - 18:00)
Channel: Track 03

FROM 2020 (Thursday, September 3)
Channel: TBA

Plenary Presentation 1 (Friday, September 4, 09:00 - 09:50)
Channel: Track Plenary

Plenary Presentation 2 (Friday, September 4, 14:00 - 14:50)
Channel: Track Plenary

Plenary Presentation 3 (Friday, September 4, 15:00 - 15:50)
Channel: Track Plenary

Computer Vision: Automated Driving 1 (Friday, September 4, 10:00 - 11:20)
Channel: Track 03

Intelligent Systems: Fuzzy Reasoning and Brain Signal Classification (Friday, September 4, 10:00 - 11:20)
Channel: Track 01

Distributed Computing and Networking: Blockchain Applications (Friday, September 4, 10:00 - 11:20)
Channel: Track 02

Computer Vision: Automated Driving 2 (Friday, September 4, 11:30 - 12:50)
Channel: Track 03

Intelligent Systems: Intelligence in Software projects (Friday, September 4, 11:30 - 12:50)
Channel: Track 01

Distributed Computing and Networking: Systems Quality and Safety (Friday, September 4, 11:30 - 12:50)
Channel: Track 02

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 1 (Friday, September 4, 16:00 - 17:20)
Channel: Track 03

Intelligent Systems: Natural Language Processing (Friday, September 4, 16:00 - 17:20)
Channel: Track 01

HiPerGRID: Modern Streaming and Infrastructure (Friday, September 4, 16:00 - 17:20)
Channel: Track 02

Computer Vision: Deep Learning Based Perception for Autonomous Systems - 2 (Friday, September 4, 17:30 - 18:50)
Channel: Track 03

Intelligent Systems: System Optimization (Friday, September 4, 17:30 - 18:50)
Channel: Track 01

HiPerGRID: Satellite Data Processing (Friday, September 4, 17:30 - 18:50)
Channel: Track 02

Intelligent Systems: Knowledge and Rule Based Systems (Saturday, September 5, 09:00 - 10:20)
Channel: Track 01

Intelligent Systems: Applications (Saturday, September 5, 09:00 - 10:20)
Channel: Track 02

Computer Vision: Technologies and Applications - 1 (Saturday, September 5, 09:00 - 10:20)
Channel: Track 03

Intelligent Systems: Learning (Saturday, September 5, 10:30 - 11:50)
Channel: Track 01

Intelligent Systems: Security and Privacy (Saturday, September 5, 10:30 - 11:50)
Channel: Track 02

Computer Vision: Technologies and Applications - 2 (Saturday, September 5, 10:30 - 11:50)
Channel: Track 03

Computer Vision: Medical Imaging - 3 (Saturday, September 5, 12:00 - 13:20)
Channel: Track 03