Two Computer Science MSc theses to be presented on 21 August

Wednesday, 21 August there will be two master thesis presentations in Computer Science at Lund University, Faculty of Engineering.

The presentations will take place in E:4130 (Lucas) and E:2116. See information for each presentation.

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:4130 (Lucas)

Presenter: Carl Wikström
Title: Evaluating synthetic data for enhancement of object detection models
Examiner: Jacek Malec
Supervisors: Volker Krueger(LTH), Patric Fröjd (Verisure Innovations)

This thesis investigates the potential of using synthetic data to enhance object detection models in security surveillance contexts. Synthetic data was generated with the Unity Perception package and evaluated using real-world security footage. The study compared these results with those from a model trained on an equivalent number of COCO dataset images. Findings indicate that while synthetic data alone did not match the performance of real data, fine-tuning the model with a small target domain dataset allowed synthetic data to perform equivalently. The research highlights the benefits of synthetic data, especially when real data is scarce, and shows promise for detecting specific behaviors, such as crawling versus standing individuals. This study contributes to the ongoing discussion about the practicality of synthetic data in surveillance applications, underscoring both the challenges and opportunities of leveraging synthetic datasets for enhancing object detection models in diverse security scenarios.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240821_11Wikstrom.pdf

11:15-12:00 in E:2116

Presenters: Emelie Bondesson, Daniel Ahlberg
Title: Comparing compression algorithms for transaction logs in graph native databases
Examiner: Alma Orucevic-Alagic
Supervisors: Jonas Skeppstedt (LTH), Nils Ceberg (Neo4j), Lukas Gustavsson (Neo4j)

In today's society, we store more data every day in different types of databases. One of these types that has become more popular in the last decade is graph databases. These databases store both the data itself and auxiliary data where logs detailing the change history for that specific database are an important part. With today's distributed systems, the storage and transfer of logs is a key function to ensure uniformity between instances of the database. With growing amounts of data, minimizing the auxiliary data is more important than ever to reduce costly memory consumption and energy usage. This thesis tested the three compression algorithms GZIP, ZSTD and LZ4 to see if it was possible to compress the transaction logs for Neo4j, a native graph database, without adding too much overhead during runtime. The efficacy of these algorithms was measured in terms of compression ratio and execution time.

Firstly, the execution flow of committing transactions was studied to understand the underlying design of the system. With this design and the ACID principles in mind, a streaming compression library was used during commits to compress the payload. The design was tested and benchmarked with Neo4j's internal suite.

The results showed that LZ4 was the fastest, with nearly no impact on execution time, while achieving a compression ratio of 97%-21% of the original size. GZIP and ZSTD achieved similar ratios to each other at 75%-10%, with a negative effect on execution time, up to 8 times slower than the baseline. The size of the transactions was the main factor affecting the compression ratio. The conclusion is that the logs contain many patterns that make them possible to compress but depending on the demands on performance or memory usage different algorithms may be best suited.

Link to popular science summary: To be uploaded