Sixteen MSc theses to be presented on Wednesday June 11, 2025

Wednesday June 11 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Sixteen theses will be presented.

You will find information about how to follow along under each presentation. There will be presentations in three different rooms: E:2116, E:2405 (Glasburen) and E:4130 (Lucas). See room for each presentation. A preliminary schedule follows.

Please note that there will also be thesis presentations on Wednesday June 4, schedule at: https://cs.lth.se/kalendarium/?evenemang=cs-msc-thesis-presentation-day-june-4-2025

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.

09:15-10:00 in E:2116

Presenters: Desirée Ohlsson, Vincent Ossler Röyter
Title: Digital Twin with Live Feeds from Network Cameras
Examiner: Flavius Gruian
Supervisors: Michael Doggett (LTH), VinayVenkanagoud Patil (Axis Communications AB)

This thesis investigates whether it is possible to update a digital twin using live video from network cameras. We developed a pipeline that detects persistent changes in the scene's background, segments the changed regions, and estimates their depth using monocular depth estimation models. Based on this data, 3D objects are reconstructed and placed at the correct geographical location in a digital twin, using known extrinsic parameters from a calibrated camera.

To evaluate the system’s performance, we compare images from a real camera with the rendered view from a virtual camera placed at the same position in the digital twin. The results show that new 3D objects can be inserted at relatively accurate depths and positions, allowing the digital twin to reflect changes in the physical scene. While the accuracy in terms of placement and visual quality is limited, primarily by the performance of the depth estimation models and the limitations of single-perspective input, the results provide a working proof of concept and a foundation for future research in this area.

Link to popular science summary: Link to be added

10:15-11:00 in E:2116

Presenter: Niklas Sandén
Title: Improving Perceptual Video Quality Metrics by Considering Object Motion
Examiner: Flavius Gruian
Supervisors: Michael Doggett (LTH), Pontus Ebelin (NVIDIA)

Quality metrics are widely used in computer graphics research to evaluate the output of different rendering algorithms. However, commonly used metrics such as PSNR and SSIM fail to capture many elements of human perception, leading to results that often do not align with human judgments. This has led to increased popularity of more perceptual metrics, such as FLIP.

ColorVideoVDP is a recent video quality metric that incorporates spatiotemporal aspects of human vision. However, one of its key limitations is that it does not model eye motion. This can lead to inaccurate assessments when evaluating videos with moving objects. This thesis proposes several extensions to the temporal component of ColorVideoVDP to address this limitation, including simulating the eye tracking moving objects by reprojecting video frames based on object motion.

We evaluate our metric using the datasets used to train ColorVideoVDP, as well as on common scenes used in computer graphics research.

Link to popular science summary: Link to be added

10:15-11:00 in E:2405 (Glasburen)

Presenters: Elise Olsson, Ebba Johannesson
Title: Evaluating re-ranking methods in LLM search
Examiner: Markus Borg
Supervisors: Marcus Klang (LTH)

This project explores re-ranking methods in the context of information retrieval, where an initial set of candidate documents is retrieved using a semantic search method. We implement and evaluate three different re-ranking models: a cross-encoder, a support vector machine (SVM) model, and a bi-encoder model. The goal is to compare their effectiveness in terms of ranking quality and processing speed. To support this comparison, we developed an evaluation framework and trained the models on multiple datasets.

Our results show that the cross-encoder consistently outperforms the other models on our product-based datasets in terms of ranking quality, while the bi-encoder model achieves the best results on our passage-based dataset, but only slightly better than the initial retrieval quality. In terms of efficiency, the SVM model is the most suitable option when using only a CPU, but the cross-encoder becomes the fastest when executed on a GPU.

Link to popular science summary: Link to be added

11:15-12:00 in E:4130 (Lucas)

Presenters: Simon Börjesson, Erik Ersmark
Title: Time series forecasting of truck data with foundation models and space-filling curves
Examiner: Elin Anna Topp (LTH)
Supervisors: Pierre Nugues (LTH), Christian Berger (GU/Chalmers), Klara Eliasson (AB Volvo)

During operation, trucks communicate internally using a CAN bus. Logging these signals over time allows for various analyzes. Additionally, time series forecasting has been used in many fields where large amounts of proprietary time series data is available. However, the potential for improving advanced driver-assistance systems by using time series forecasting with foundation models and the potential of using space-filling curves for dimensionality reduction on this data has not been explored. Here we show that the performance of foundation models is promising, but that space-filling curves likely cannot be used directly with foundation models. We found the performance of several models commendable when predicting the steering wheel angle of different sections of the latter half of a roundabout maneuver, especially when fine-tuning, and Chronos-Bolt was the best-performing model in the cohort. We believe this shows potential for fine-tuning foundation models for general CAN data, not only from specific scenarios.

Link to popular science summary: Link to be added

13:15-14:00 in E:2116

Presenter: Anna Johannesson
Title: Solving Non-Linear PDEs Using CUDA
Examiner: Michael Doggett
Supervisors: Flavius Gruian (LTH), Michail Boulasikis (LTH)

Partial differential equations (PDEs) are central in many scientific research areas, including modeling physical phenomena and economic forecasting. While the general solutions of PDEs can be complex, we can determine a particular solution by introducing boundary conditions and/or initial conditions. This is called an Initial Boundary Value Problem (IBVP).

Thalassa is a framework that generates solvers for IBVPs in PyTorch, targeting both CPU and GPU execution. One goal of this thesis was to extend Thalassa to produce solvers in CUDA C++. We compared the performance of the PyTorch solvers to their CUDA C++ counterparts, using different PDEs and problem sizes, focusing on the execution time. We also profiled the CUDA C++ solvers in order to identify bottlenecks that would affect the speed of the CUDA C++ solvers.

Our results show that the CUDA C++ solvers outperformed their equivalent PyTorch solvers, targeting both the CPU and the GPU. Speedups ranged from 25x to 558x compared to PyTorch CPU solvers, and from 20x to 596x compared to PyTorch GPU solvers.

Link to popular science summary: Link to be added

13:15-14:00 in E:2405 (Glasburen)

Presenters: Joel Pistora, Gustav Kristiansson
Title: Enhancing the Prioritisation of Security Requirements in Software Development: A Cost-Effectiveness Perspective
Examiner: Björn Regnell
Supervisors: Alma Orucevic-Alagic (LTH), Gustav Lundsgård (IKEA)

A major issue in cybersecurity is the lack of relevant knowledge for prioritising security requirements due to the diverse context of software systems. A cross-team priority order if often ineffective, highlighting the need for a method that accounts for system-specific context parameters when prioritising security requirements.

This study was conducted at a large-scale organisation, with the goal to investigate how to formulate a contextual requirement prioritisation method. Through a workshop involving cybersecurity experts, key system parameters that influence prioritisation decisions were identified. These parameters were used as context for a survey in which respondents made pairwise comparisons of security requirements. The collected data was analysed using a developed scoring model.

The results indicate that prioritisation changes based on system context and that respondents reached general consensus. Based on these findings, a six-step method was proposed to support the contextual prioritisation of security requirements in practice.

Link to popular science summary: Link to be added

13:15-14:00 in E:4130 (Lucas)

Presenters: Albin Green, Clemens Christierson
Title: Applications of ML-based Methods in Live Stream Video Compression
Examiner: Jacek Malec
Supervisor: Simon Kristoffersson Lind (LTH)

Given the vast amount of video content transmitted over networks, demand for efficient video compression is high. Recent work in AI-based video compression has shown promising results in terms of quality, but rarely in terms of real time performance. This thesis presents a video encoder and decoder that leverages machine learning-based motion estimation and super resolution, with the goal of real time video encoding. Its performance is evaluated alongside H.264 across a varied dataset, analyzing compression efficiency, visual quality, and computational cost.

Different metrics for computational efficiency, compression ratio and visual quality were used to determine the viability of these AI-based methods in comparison with H.264. Results show that both temporal and spatial video complexity influence compression ratio and visual quality. Further, the method shows promising performance for real time scenarios, and avoids compression artifacts common in traditionally encoded video.

Link to popular science summary: Link to be added

14:15-15:00 in E:2116

Presenters: Anna Grimlund, Gustav Lindqvist
Title: Learned vs. Traditional Indexing of Geo-spatial data: A Comparative Analysis
Examiner: Flavius Gruian
Supervisors: Gareth Callanan (LTH), Susanna Rezende (LTH), Per Svensson (Företaget)

This thesis presents a comparative analysis of learned versus traditional index structures for handling large-scale geospatial data. Given the increasing volume and importance of geographical data in modern applications, efficient indexing is crucial for optimizing query performance. The objective was to compare learned index structures to established traditional methods focusing on lookup times for point, window, and kNN queries, as well as space cost, build time and scalability. The study implemented and tested five indexing techniques, including the traditional R-tree and KD-tree, alongside the learned ZM-index, IF-X index, and Flood, using OpenStreetMap datasets in both in-memory and on-disk scenarios. Results indicate that while no single learned index consistently outperformed across all tests, they generally demonstrated superior scalability and significantly lower memory footprints compared to traditional tree-based structures. The findings highlight the potential of learned indexes for improving efficiency in geospatial data management, while also identifying potential limitations.

Link to popular science summary: Link to be added

14:15-15:00 in E:2405 (Glasburen)

Presenters: David Andrésen, Torsten Sandell
Title: Relational Object Tracking in Multi-Camera Systems
Examiner: Volker Krueger
Supervisors: Simon Kristoffersson Lind (LTH), Panagiotis Bitharis (Axis Communications AB)

The goal of this thesis is to track objects in multi-camera views as well as relations between the objects based on metadata with geographical coordinates output from Axis’ cameras. This tracking may assist in finding when and where objects have interacted with each other. We implemented a particle filter which filtered the objects’ coordinates and cross-referenced with other cameras’ objects for potential matches. The implementation was designed to work both in overlapping and non-overlapping camera views, and our application showed adequate performance in both cases.

When objects have something in common, for instance, when a face belongs to a human, we call this a relationship. We track relations by matching overlapping bounding boxes for humans and faces, and by observing the location of humans to match them with other humans. This is trivial to extend to other objects such as bags or cars as well.

Link to popular science summary: Link to be added

14:15-15:00 in E:4130 (Lucas)

Presenters: Salam Jonasson, Albin Andersson
Title: Modeling Loudspeaker Nonlinearities with Deep Learning
Examiner: Jacek Malec
Supervisor: Pierre Nugues (LTH)

Loudspeakers inherently introduce both linear and nonlinear transformations to the sound they produce. While standard lightweight echo cancellers can effectively adapt to and remove the linearly transformed echo, they struggle to handle the nonlinearities unless they are explicitly modeled.

This work aims to model and predict such nonlinearities produced by the loudspeaker and perceived by its microphone, employing advanced machine learning techniques. The objective is to overcome a key limitation of standard echo cancellation algorithms: their inability to remove the contorted portion of the echo, thereby improving overall echo cancellation performance. Furthermore, this thesis benchmarks the performance of different neural network architectures in accurately modeling these nonlinearities.

The results demonstrate that it is possible to effectively model loudspeaker nonlinearities, leading to significant improvements in echo cancellation performance compared to traditional linear methods. Adjusting the signal through preprocessing to account for linear transformations before inputting it into the machine learning model enables the model to concentrate mainly on nonlinear elements. This approach leads to improved accuracy in prediction and cancellation of the nonlinear components.

Link to popular science summary: Link to be added

15:15-16:00 in E:2116

Presenters: Andreas Bergqvist, Karl-Philip Blé Cato
Title: Exploring the extended software bill of materials
Examiner: Emelie Engström
Supervisors: Lars Bendix (LTH), Christer Wibom (ABB AB), Jonas Stigeberg (ABB AB)

In recent years, there has been an increase in attacks on the software supply chain, sparking interest in the software bill of materials (SBoM). Adopters hope that SBoM will aid in vulnerability analysis. This study explores how early implementation of SBoM can improve the software development process and identifies key data points.

We conducted interviews at ABB to determine use cases and data points for an eSBoM (extended Software Bill of Materials). Our contributions are defining a standard SBoM, fulfilling the requirements set by CRA and Executive order 14028, archetypes grouping use cases by function and purpose, a catalog of use cases and data points of interest for ABB, and a proof of concept for a subset of these use cases. Based on interview material and the potential for generating SBoMs during the pipeline build, we see potential benefits of early SBoM implementation.

Link to popular science summary: Link to be added

15:15-16:00 in E:2405 (Glasburen)

Presenters: Marcus Fröberg, Måns Englund
Title: Exploring the Interplay Between Code Smells and Energy Consumption
Examiner: Christoph Reichenbach
Supervisors: Markus Borg (LTH), Adam Tornhill (CodeScene AB)

As energy consumption becomes a growing concern in software development, understanding how code quality influences energy usage is increasingly important. This study investigates the relationship between CodeHealth, code smells, and energy consumption. We conduct a two-part empirical study: a correlation analysis between code quality metrics and energy usage, followed by a refactoring experiment. While past research has focused on performance-based or Android-specific smells, our work explores a broader range, addressing a gap in current research. We aim to determine whether software quality metrics correlate with energy consumption and whether refactoring specific smells has an impact. Results show a moderate correlation between CodeScene’s CodeHealth metric and energy usage, and a strong correlation between LOC and both CodeHealth and number of smells. Regarding our second experiment, results were inconclusive due to measurement difficulties. However, our general findings support the view that improving code quality can enhance software energy efficiency.

Link to popular science summary: Link to be added

15:15-16:00 in E:4130 (Lucas)

Presenters: Georgios Chavales, Edoardo Giorgio Piotr Vaira
Title: Transforming Project Documentation with AI
Examiner: Jacek Malek
Supervisors: Pierre Nugues (LTH), Nourhan Aloush (Robert Bosch Aktiebolag)

Rapidly expanding project documentation in Agile and DevOps environments presents significant challenges for stakeholders needing to efficiently access and understand system information. Manual searching is time-consuming and error-prone, while existing tools often fall short. This thesis introduces an end-to-end retrieval-augmented generation (RAG) chatbot, developed with Bosch, using GitHub repositories as knowledge sources. It automates ingestion of heterogeneous project artifacts, indexes them in a hybrid dense–sparse vector store and supports natural-language queries. Our flexible ingestion pipeline handles diverse formats and incremental updates, while the hybrid retriever, with optional pseudo‐query augmentation, efficiently brings up relevant information. An advanced generative module synthesizes concise, context-aware answers, while session-level utilities enable the automated synthesizing of documentation. Quantitative and qualitative evaluations demonstrate significant performance in retrieval accuracy and reduced documentation search time, highlighting the system’s potential to streamline industrial documentation workflows. Finally, we integrated the architecture in a web application and deployed it on Azure.

Link to popular science summary: Link to be added

16:15-17:00 in E:2116

Presenter: Marcel Urrutia Nilsson
Title: Predicting Psychological Scale Scores and Deltas from Structured Word Responses: A Comparative Study of Regression Pipelines
Examiner: Maj Stenmark
Supervisor: Dennis Medved (LTH)

Background: Despite the potential benefits of using structured word responses for predicting psychological construct using rating scale scores, there is a lack of research in this area, particularly in the context of construct scale delta prediction. Thus, a comparative study of regression pipelines could provide valuable insights into the most effective method for this type of analysis.

Aim: This thesis aims to investigate the effectiveness of various regression pipelines in predicting psychological construct scale scores and deltas using structured word responses.

Method: The data used is from an earlier study with 477 participants who completed the Harmony in Life Scale (HILS) and the Satisfaction With Life Scale (SWLS) twice. Word embedding extraction and PCA were used for preprocessing, and two feature engineering strategies were employed for delta prediction, Delta vector and T1-score. Hyperparameters were optimized using Bayesian optimization, and four regression models were compared: ridge regression, Bayesian ridge regression, random forest, and using the package AutoKeras StructuredDataRegressor to create deep neural network models.

Results: AutoKeras outperformed all other models for both score and delta prediction, with Pearson r-values of 0.929 and 0.918 for the HILS and the SWLS score prediction, and 0.933 and 0.874 for the HILS and the SWLS delta prediction. No other model achieved an r-value higher than 0.8 in any task.

Conclusions: The findings suggest that AutoKeras is a promising approach for predicting psychological construct scale scores and deltas from structured word responses. However, several modifications are necessary to ensure the validity and comparability of the results, including addressing data leakage and adopting a format more consistent with earlier studies. Further research and bigger datasets is needed to evaluate the generalizability of these promising results to other datasets and scales.

Link to popular science summary: To be added

16:15-17:00 in E:2405 (Glasburen)

Presenter: Adam Ohlsson
Title: How does refactoring affect performance?
Examiner: Niklas Fors
Supervisor: Christoph Reichenbach (LTH)

Refactoring is a recognized technique for improving code structure. Developers apply refactoring to improve quality and performance aspects of code. However, the costs of Java refactoring, and the trade-offs that we make between quality and performance in our low-level designs, are not well understood. In this work, we use a semi-automated approach to explore how Java refactoring affects performance by evaluating speedup effects of 12 common Java refactoring types on 5 curated benchmarks in four compiler and runtime configurations across two hardware settings. We found that the evaluated refactoring types do not affect speedup on average, but we believe there are instances where individual refactorings have a significant impact.

Link to popular science summary: To be added

16:15-17:00 in E:4130 (Lucas)

Presenters: Simon Hallefält, Olof Ekenberg
Title: Crafting a Heterogeneous Multi-Agent Pathfinding Simulation
Examiner: Jacek Malec
Supervisor: Volker Krueger (LTH)

Heterogeneous multi-agent pathfinding (H-MAPF) is a generalized form of multi-agent pathfinding where agents can have differing sizes, speeds, or other characteristics. This paper discusses the implementation of a Java program to simulate H-MAPF in a warehouse setting. It covers the design decisions made, potential applications of the program, and its limitations, as well as providing an overview of the program and how to run it.

Link to popular science summary: To be added