Events
CS MSc Thesis Presentations 16 June 2023
Föreläsning
From:
2023-06-16 11:15
to
16:00
Place: E:2116
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se
Save event to your calendar
Four Computer Science MSc theses to be presented on 16 June
Friday, 16 June there will be four master thesis presentations in Computer Science at Lund University, Faculty of Engineering.
The presentations will take place in room E:2116.
Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.
11:15-12:00 in E:2116
Presenter: Axel Leander
Title: Using Transformers to Extract Data From Receipts
Examiner: Flavius Gruian
Supervisor: Pierre Nugues (LTH)
Data Extraction is a time consuming task if done manually. For computers, automating this process can be achieved through various methods. This thesis compares a regular expression system which extracts date-expressions from receipts to a machine learning system with the transformer architecture. The idea is that a machine learning system has the possibility of being more robust to spelling errors and OCR-misses. To train the machine learning models I produced training data from the regular expression system which it was compared to as well as generating fake data. The final performance of the machine learning system was similar to that of the regex module which it was compared to. The machine learning system also showed indications of ignoring minor OCR-misses and the ability to extract dates on formats with minor variations from those which it was trained on.
Link to popular science summary: To be uploaded
13:15-14:00 in E:2116 (presentation was added to this page)
Presenters: Leo Westerberg, Simon Erlandsson
Title: Compression Algorithms for Geometries Supporting Operations
Examiner: Per Andersson
Supervisors: Jonas Skeppstedt (LTH), Hampus Londögård (AFRY)
Maps-service providers use vast amounts of geometric data to represent the world’s structure. With growing amounts of data, the required storage and transmission capacities increase. Existing compression algorithms can reduce the data size on the cost of operability. Namely, any operation requires the entire compressed geometry to be unpacked. This thesis explores the possibilities of creating a compression format that reduces geometries' data size while maintaining speed on some specific operations. The implemented format utilizes delta encoding, a map-specific coordinate structure, and entropy encoding for size reduction. Coordinates are segmented into blocks for partial decompression, avoiding unnecessary decoding of non-relevant geometry sections. For instance, calculating intersections between shapes only requires overlapping blocks. Testing revealed a 59% reduction in size and a 60% speed increase compared to the baseline when calculating intersections over large geometries. Partial decompression of geometries is largely unexplored in academia, but the results of the thesis indicate that the area may be of interest to investigate further.
Link to popular science summary: To be uploaded
14:15-15:00 in E:2116
Presenters: Kristina Patrikson, Gustav Klotz
Title: Improving artificial validation data using scene analysis
Examiner: Flavius Gruian
Supervisors: Michael Doggett (LTH), Mikael Murstam (Axis Communications AB), Linus Jacobson (Axis Communications AB)
This project explores the possibility of creating virtual validation data by using augmented reality. We have developed a shader that aims to make a virtual car look as similar as possible to a real car, especially in low-light situations. The shader takes the output from a graph cut-based shadow segmentation algorithm in order to match the lighting of the scene. A virtual car equipped with this shader is compared to a realistic model car both by using output values from a vehicle-detecting AI and by visual assessment. The results are also compared to another virtual car using a simpler shader that does not take the scene into account. Our results show that our shader makes the virtual car look visually similar to the model car compared to the simpler shader, but it was not possible to show that the AI found our shader more realistic than the simpler shader.
Link to popular science summary: To be uploaded
15:15-16:00 in E:2116 (presentation was added to this page)
Presenter: Filip Jergle Almquist
Title: Understanding the suitability of linear algebra algorithms for a generic parallel computing API
Examiner: Flavius Gruian
Supervisors: Jonas Skeppstedt (LTH), Anders Åhlander (Saab AB)
This Master's thesis evaluates oneAPI, a unified parallel computing API, for its application in radar signal processing within active electronically scanned array (AESA) radar systems. These systems necessitate real-time data processing, sometimes exceeding conventional CPU capabilities, leading to the need for code reusability across varied parallel hardware like GPUs and FPGAs. This study, comparing oneAPI against established alternatives, evaluates its performance, portability, and productivity. Findings show that despite its novelty, oneAPI is competitive, even superior in some cases, albeit lagging behind native CUDA on Nvidia GPUs. oneAPI also exhibits high code reusability across hardware, suggesting potential benefits for code maintainability, cost efficiency, and product lifespan. However, limitations exist due to oneAPI's immaturity and further exploration of more diverse data parallel algorithms and additional hardware types is needed for a complete understanding of oneAPI's potential in high-performance computing.
Link to popular science summary: To be uploaded