Захист бакалаврських дипломних робіт, 2020

9-13 червня 2020 року на факультеті прикладних наук УКУ відбудуться захисти дипломних робіт студентів бакалаврської програми з комп’ютерних наук. Подія відбуватиметься онлайн.

Розклад проведення захистів

9 червня

Петрук Мар’ян Олегович, Face-reenactment: generation flexibility and identity preservation 

Abstract.   Face-reenactment also knows as puppetry, has become very popular in recent years. The proposed task requires generating new face expression while preserving person identity and scene features. In this work, we propose advancements forecent novel methods of accurate face-reenactment synthesis. We present results using a flexible generation module, and compare different families of encoding backbones, introduce identity loss to preserve a person’s identity in image generation with state-of-the-art models in the deep face-recognition domain. We also provide improvements in the training procedure and test on approach weaknesses.

Захарченко Ірина Русланівна, Basketball Pose-based Action Recognition 

Abstract.  Action detection on a team sport is a challenging task, while sports analysis is on-demand and in high interest. A great number of researchers try to make analysis automated. Despite enormous success in image classification using deep learning, action recognition in the video remains a difficult task, and at present no good solution exists in terms of accuracy and speed. The main challenge in action recognition is to design architecture that will capture both spatial and temporal information. In team sports action analysis, the serious challenges are that we have multiple players performing simultaneously different actions, the players are constantly moving, there are occlusions, the camera itself is moving. The proposed method is able to simultaneously recognize the actions of multiple players using pose estimation, tracking, and LSTM for action classification.

Вей Роман Андрійович, NoGAN: Deblurring Images without Adversarial Training

Abstract. In this paper, we systematically study generative adversarial networks (GANs) for single image motion deblurring. Firstly, we compare adversarial loss functions, discriminators, and other training configurations to find optimal setup. Secondly, we train the blurred image classification model and use it as a pretrained discriminator in the GAN setup. We train GANs in two ways: with frozen and unfrozen discriminator weights.

Борківський Богдан Петрович, Rubik’s Cube layout construction and validation algorithm

Abstract. Many sports have some monitoring systems, that help judges to make wise decision in controversial situation, such as VAR in football. VAR allows judge to check situation from different views, so nothing is missed. In this work we propose a prototype of such system for speedcubing, i.e. sport of solving different puzzles, such as Rubik’s cube on time. This system can help to prevent mistakes on scrambling stage, when cube is mixed from solved state follow- ing the specific algorithm. The developed system is able to locate cube on scene and based on observable parts can check layout with a list of allowed layouts.

Забульський Володимир Васильович, Reflection Removal with Generative Adversarial Networks

Abstract. We propose an approach for the single image reflection removal problem. Our model is based on a feature pyramid network (FPN), trained with adversarial and perceptual losses. Additionally, we address the problem of bright spots removal when only a small portion of an image is covered with the reflection. The difficulty of collecting real-world data makes the problem even harder to solve. We propose a novel method of collecting real-world data, that does not require any additional devices but a camera, and is cheaper than the existing ones. We collected a small dataset with this approach.

Темник Мар’яна Любомирівна, Lane segmentation for advanced driver-assistance systems

Abstract.  Lane detection is an image analysis problem of detecting lane markings on the road. Recently it became one of the crucial parts of most Advanced diver-assistance systems. With the current progress in deep learning, plenty of new methods have been developed for lane detection. However, even novel powerful CNN architectures are struggling to give a satisfactory result along with acceptable performance. Recent studies show that developing a fast, robust and accurate model for real-time lane detection is a hard task due to external factors such as weather conditions, lighting, traffic, complex marking shapes, thin or invisible lanes on rough and hilly roads, etc.

Борсук Василь Юліанович, HRGAN: High-Resolution Representation Learning for Image Deblurring

Abstract.  Existing state-of-the-art single-image motion deblurring frameworks first encode the input image as a semantically rich low-resolution representation and then recover the high-resolution image from this representation. The network which generates an image from low-resolution features is not the right choice for pixel-level predictions as it fails to recover finer texture details. We introduce the idea of the High-Resolution Network (HRNet) to image deblurring. It maintains high-resolution representation through the whole process instead of recovering high resolution from low resolution. Additionally, we propose to use the CutMix augmentation strategy to enhance the performance of our network


10 червня

Косаревич Іван Ростиславович, Face reenactment with GANs using landmark representation of a face

Abstract.  Face reenactment is an emerging technology that attracts high interest in recent years. It aims at generating face with the identity of one person (known as target) and facial expression from another (source). Many existing methods are limited to reenact a predefined personality of either source or target. In this study, we present the approach that is agnostic to the identity of source and target and observes only a single image of each of them. Our method is based on recently introduced Generative adversarial networks (GANs). We experimentally find a proper GAN loss for our system. An accurate expression transfer from a source person is essential forface reenactment. In this study, we examine different approaches to achieve it and design a landmark loss function based on our novel landmark detector.

Гірна Марія Святославівна, End2end image analysis of single-cell gel electrophoresis

Abstract.  Single-cell gel electrophoresis is the standard test used by biomedical researchers to analyze damage to the cell. Currently, this test is only done using standard image processing techniques, that skews the outputs, requires manual work and/or human supervision. Other problems with current solutions include poor usability, lack of flexibility, and high price for commercial applications. In this work, we create a deep learning-based end2end pipeline, that receives images from the test as an input, and produces damage metrics as an output. We have trained UNet with SE-ResNet50 encoder on the custom-created synthetic dataset, which achieves the dice coefficient of 76.8. We hope that the results of this work will become the base of the easy-to-use open-source application available for any researcher.

Костишин Ірина Ярославівна, On estimation methods of image visual aesthetics

Abstract.  Evaluation of image aesthetics has been a longstanding problem in image processing and computer vision. Nowadays this topic is discussed more than ever, researchers attempt not only to measure the aesthetic quality of images but also to find new applications of this methods. In this work, we research image assessment methods, their pitfalls. We try to propose new assessment methods for these tasks using activation maps and explore the possibility of using such neural networks for training Image Restoration GANs. This work also incorporates the study of the relevance of existing methods on the example of their work on data not specially prepared for this specific task.

Лєпєшов Костянтин Сергійович, Manhattan Frame Detection in Lens Distorted Images Using Lines of Circle Centers

Abstract. Camera auto-calibration from a single image with radial distortion is a prevalent task in computer vision. Most of the existing approaches are based on the same process of extraction of features, such as circles, from the image. Since those features are noisy, the error is propagated to the higher level, and the final estimations are incorrect. We try to incorporate the constraints created under the assumption of the division model radial distortion to create a simple approach that gives soft estimates of three Manhattan directions. For this problem, we adapt a well-known Expectation Maximisation algorithm. We combine it with different initialization and filtering steps that we form based on the division model and Manhattan world assumptions. We show the performance of the proposed approach on YORK Urban Database (YUD) and AIT Dataset of inside and outside scenes. Besides, we experiment with the proposed initializations and filtering procedures.

Турій Борис Олегович, Development of a car information retrieval system by license plate

Abstract. During four years of study, we were able to work on almost all available topics in Computer Science. Programming, Algorithms, Robotics, Operating Systems, Artificial Intelligence, Networks, Security, Databases, Cloud Computing, Web Development, and many more. This work is an essence of courses that interested me the most – my trial and final exam. Here Databases, Programming, Cloud Engineering, Web Development will combine into powerful architecture – that will be easy to reproduce and scale. There will be implemented – gathering, analyzing, storing, and providing car data.

Шпот Наталя-Яна Миколаївна, Accuracy and Bias of selfie detection on open data

Abstract. There are many challenges related to the openness of the Wikimedia Commons image upload platform, and one of them is about making sure to get high-quality content in. Goes without saying, selfies are not precisely the ideal wanted content for a platform whose aim is to represent the world’s knowledge through pictorial representations. One way to automatically check the data quality in the domain of computer vision is to design a selfie detector that, given an image, can automatically predict whether it is a selfie or not. Thus in this thesis, we are using state-of-the-art models to create a classifier that, given an image, can say whether the image is a selfie, a person, or neither of that. With such a classifier, it would be easier to automatically detect and scale selfies for Wikimedia or other platforms that have humans in the loop to check the quality of user-generated content. In addition to this we examine whether approaches of our choice show bias in demographics such as race, gender, and age. Furthermore, we will introduce two datasets for our project: one containing selfies, pictures with persons and random pictures, and another containing a smaller set of pictures of persons along with the demographic metadata.

Баценко Тетяна Максимівна, Development of a system for monitoring blood donor searches in Social Networks

Abstract. Blood donor shortages are happening every day, and it can cost people’s lives. The outbreak of Covid-19 has caused even more severe blood shortages, as people and organizations cancel planned donations because of the quarantine. It is important to understand how and where Twitter is used to find blood donors quickly. Social media monitoring is being used widely by businesses to capture trends and understand end-users, why not use it for good too. This work provides an extensive summary of what has been done before, as well as presenting a set of features and tools that can be used to identify blood donor requests on Twitter. The result of this work is a real-time monitoring system for blood donor requests.

Юзьків Вікторія Ігорівна, Judge a book by its cover. Generative approach.

Abstract.  This work is an example of generative art and design application in publication graphic design. It shows how programming may be applied in a book covers generation. The work combines such fields of Computer Science as Natural Language Processing (processing text data to get its structure and features), Algorithms and Data Structures (representing and working with a tree data structure), and Data Visualization (generation of cover elements). All covers are generated for Artemis Fowl – a series of novels by Eoin Colfer.


11 червня

Антентик Юрій Миколайович, Posteriograms Postprocessing for Multi-Pitch Estimation

Аstract.  Music Transcription is a task of converting a musical recording into sheet music for further reproduction. The problem is still unsolved and requires a high level of expertise. Most of the works split the task into several subproblems. First of them is called frame-level transcription, which predicts the set of fundamental frequencies in the original recording for every frame. This subproblem is the main focus of this work. The solution to frame-level transcription is called a piano-roll representation – a binary matrix which represents whether the given note has been played in the frame or not. However, most of the approaches do not produce a piano-roll representation in an end-to-end fashion. They rather output a posteriogram – real matrix with the same dimensions, which represents the level of uncertainty of whether note has been played during the frame. Ycart and Benetos, 2018 shows that Long Short Term memory network can be trained to post-process the posteriograms and improve the piano-roll representation instead of simply cropping the posteriogram at some value. In this work, we train more robust LSTM network and experiment with different types of posteriograms.

Боровець Андрій Андрійович, Development of a visualization system for neural network results on medical images

Abstract. Heart stenosis corresponds to a frequent problem with the heart arteries when the valve is getting too narrow and does not open properly. The flaps of a valve may become thicker or bound together, and as a result, preventing the valve from fully opening. This problem can be detected and treated, if diagnosed in time. To determine if patients’ conditions are critical or not – may require a significant amount of time of the doctor. Diagnosis timing is one of the most critical factors with this particular disease, frequently determining the outcome for the patient. Considering the current general medical service practice, there is no apparent way to create a prioritized patients list, so that patients with more critical or urgent conditions could be treated first.

Журавчак Андрій Богданович, Human Activity Recognition based on WiFi CSI data

Єжов Віктор Олександрович, Android application for anomaly detection in car engine system

Abstract. Nowadays smartphones are playing a major role in our everyday life. The computational power of flagman versions can be compared with some of the laptops. Having all that power in our pockets gives us an opportunity to create apps with capabilities to analyze large amount of data, perform computations for making predictions, and in the end, increase life comfort of the user. In this work, I would like to describe the process of creating Android application, which aims to detect anomalies in car engine system.

Качмар Павло Олегович, Artificial intelligence in sports: trends analysis and forecast based on news articles

Abstract.  This thesis focuses on analysing trends at the intersection of artificial intelligence and the sports industry. News articles from the past seven years are the primary source of data for this analysis. With the use of natural language processing, keywords that best represent each article will be determined. Afterward, these keywords will be used to analyse current and past trends and predict future changes.

Сумуссу Марк Коджові, Basic Counter-Strike: Global Offensive demo viewer

Теплюх Юр Зеновійович, Template-based color correction in alternating lighting conditions

Abstract.   Having the right colors is one of the best ways for picture to look good. But not only humans’ eye will appreciate correctly colored image, but also a computer, especially when the task is to track object by its color. While this might not seem like a difficult one, but when you have really long-term tracking, in an environment where lightning conditions change over time, you want to be sure that your algorithm is still able to find desired object. Or maybe you cannot or don’t want to manually adjust camera color settings, but still need to track same object during the day, sunset and in the afternoon, when everything changes it colors to red and blue afterwards. The main idea behind this work is to create an automatic pipeline which will output either corrected image, or the mask, which should be applied to produce corrected image, depending on selected method. This is achieved by using a template with reference color values, which is then compared with the colors on the input image and proceeded to color estimation algorithms. Different types of templates and color estimation algorithms were compared to select the most effective to provide color stability. We show that you don’t need a special templates to achieve good results in color calibration.

Романюк Мирослава Петрівна, Improving Code Completion in Pharo Using N-gram Language Models

Abstract. Code completion is one of the essential features of any IDE and it significantly improves the developer experience and productivity. Good code completion can both speed up the development process, as well as aid the developer in API exploration. On the other hand, having a slow or inaccurate completion can be very cumbersome. Thus, it is important to find a way to make it as effective as possible. The current implementation of code completion in the Pharo IDE is based on the abstract syntax tree (AST) of source code. The AST allows us to learn the semantic role and the kind of tokens (e.g. class name, method name, literal, et cetera) and to suggest contextually relevant completions. However, the current implementation has no efficient way to sort the completion candidates, which means that sometimes the user must scroll through a long list of proposed completions to find the one that they need.


12 червня

Антонишин Юлія Миколаївна, 4-wheel stroller with smart emergency brakes system

Стасінчук Юрій Володимирович, Team of multiple autonomous UAVs interacting with static objects

Abstract.   The Mohamed Bin Zayed International Robotics Challenge (MBZIRC) is a prestigious competition, aimed at furthering the state-of-the-art in the field of autonomous robotics. This thesis presents my part of the solution to one of the tasks in the MBZIRC 2020 competition, which landed us a second place in Challenge 1 and a first place in the Grand Challenge of the competition. Specifically, this thesis focuses on the balloon popping task, which had to be solved quickly and robustly in order to compete with systems from other expert robotic teams from the whole world. In this task, a team of cooperating Unmanned Aerial Vehicles (UAVs) had to autonomously search and destroy balloons, placed in a designated arena, as fast as possible. A software and hardware overview of the used UAV platform is presented and the detection, estimation and planning algorithms of our winning solution are described in detail. Evaluation of the described methods on data from the competition is presented.

Борківський Антон Петрович, Development of an unmanned aircraft traffic management system

Abstract.  As drones become more and more popular special rules for their regulation should be created, and special systems for their management should be developed. For that purpose, SESAR has developed a list of services and strategy of their implementation to ensure safe integration of unmanned aviation into airspace. The object of our work is to develop a decomposed system that would match the SESAR list while being flexible and easy to extend. To design the architecture, we were following the IDesign framework. As a result, we provide a system with API documentation to communicate with it.

Зубрицький Маркіян Тарасович, Development of bicycle smart lock system

Поляков Михайло Хельгович, Evolution of digital organisms in truly two-dimensional memory space

Ковальчук Богдан Валентинович, Authorization server with OAuth support and recommended usage patterns

Abstract. This thesis covers the history of digital authentication and authorization and considers the main changes in different versions of the OAuth protocol. It also describes the implementation details of custom OAuth authorization servers and clients that demonstrate usage examples of different authorization grant types.

Комаренський Максим Андрійович, Street Art AR

Abstract. Augmented Reality is a new way of world perception. In the next 10 or 20 years, AR will make a revolution in every world market: Education, Advertising, Videogames, Military, Healthcare, Engineering, etc. But I do not want to wait all these years I want to start the future now.

Тріска Лука Марко Львович, Flutter app for picking and ordering tiles for interior design using AR


13 червня

Бевз Адріян Миколайович, Accessibility features and product labels for video game developers and publishers

Abstract.  Pong, which was the first-ever successful video game, was released in 1972 – almost 50 years ago. However, only for a few last years has there been a public interest in making games accessible for people with disabilities, illnesses, and injuries. Only a fraction of developers considers accessibility when designing their games. The goal of this research is to point out common accessibility problems and solutions to them. These points will be enforced by interviews with experts in the field and users. After that, we will be able to categorize the most critical accessibility features and develop a set of labels for publishers to use on their titles to let buyers know if they will be able to play the game they want to purchase comfortably.

Волоський Максим Михайлович, Development of web audio app “Music equalizer”

Abstract. The purpose of this bachelor’s thesis is to analyze the users‘ experience in the creation of the modern music and to build own web application that will help beginners or amateur musicians to learn how to use digital audio workstation. It intends to help users understand music deeply and grow up their skills in the composition of contemporary music.

Козак Олександр Андрійович, Analysis of illustration influence on human-computer Interaction

Abstract.  This paper presents a study about how digital illustration used in interface design influences human-computer interaction. An illustration is a visual interpretation of a particular concept, text or process. In this study, the name illustration is used for all its types (icons, spot illustrations, etc.) People now interact with computers differently than they did in the late 1970s when the prototype for a GUI was developed. With the development of technologies such as the Internet or a smartphone, designers moved their focus on creating products that are just usable. British tech philosopher Tom Chatfield explains (The Most Intimate Relationship in Your Life: Your Smartphone 2015) that smartphones become the most intimate relationship in lives that most people have now. People now seek for experiences that are emotional, memorable and pleasurable. Designers use a variety of tools to achieve those experiences. In particular, they tend to use an illustration. According to the statistics (Finances Online), 37% of most frequently used visual assets are original graphics, like an illustration. Apart from that, business owners also evaluated the importance of creative graphics. Companies that encourage creativity achieve 1.5 times more significant market share (Adobe). Therefore, focusing only on usability is not enough to provide meaningful and relevant experiences to users. The findings of this study are dedicated to figuring out what tasks can be covered with illustrations in UI design, how illustration affects HCI and what value it can give to the products that are created by IT specialists, in particular, computer scientists.

Мазуркевич Тарас Андрійович, Development of web application for uptime status monitoring

Abstract. For a software company, it is crucial to provide a reliable service. If something fails, engineers need to know about it as soon as possible to be able to avoid unacceptable system downtime and customers disappointment. The goal of this work was to design and develop an uptime monitoring system which would include both automated uptime monitoring and a status page. The result of this work became a part of uptime monitoring setup in a software company in which the author works. We analysed multiple uptime monitoring services and created a specification according to the company needs. It includes HTTP, WebSocket, cron jobs monitoring and web applications monitoring by simulating users actions. We implemented all the planned components. HTTP monitoring has a high level of false alarms and thus is not a fit for complex system monitoring. Passive checks have no false alarms and are a good choice for monitoring of regular jobs. We are about to start using WebSocket and web applications monitoring.

Мельничук Владислав Зіновійович, Landscape generation using procedural generation techniques

Abstract.  This work is about procedural content generation and its applications in video games. Generating a landscape is one of the ways to use procedural generation in games. The goal of this work is to test different techniques and approaches to develop a foundation for a game that is capable of creating beautiful, realistically looking landscapes. The predefined rules must fully control the generation process. The user input for the generation will be limited to the seed that defines the initial state of the generation, the parameters that control the generation, textures and 3d models. The rest of the work is automated and requires no human interaction. The results of this work can be used as a foundation for different types of games.

Степанюк Роман Андрійович, Development of analytical matching system for waste generators and disposers based on aggregated open data

Abstract. The purpose of this bachelor’s thesis is to create a platform to improve waste management in Ukraine. The service has two main functionalities: the analytics, with all the available aggregated information on waste transactions and its analysis, and B2B marketplace for owners of significant amounts of waste and recycling companies to find the best deals and start new productive partnerships.

Штогрінець Богдан Володимирович, Development of the web-application for online booking

Abstract. This paper covers user experience research about how people are using online tools for booking an appointment online, their experiences and problems. Considering the information collected and UX-design best practises – developing a web application that will cover potential users‘ need and also bring them high-quality user-experience.