Two Computer Science MSc theses to be presented on 30 January

Thursday, 30 January 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). 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.

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

Presenter: Nathalie Tiet
Title: Image Captioning Using Multimodal LLMs
Examiner: Jacek Malec
Supervisors: Pierre Nugues (LTH), Fábio Rezende (Sinch)

Image captioning has advanced significantly due to recent advancements in large language models (LLMs). By integrating LLMs with other modalities, multimodal LLMs generate richer and more descriptive image captions. However, depending on the model architectures and training sets, the captions differ in detail, length, and style, making them suitable for different use cases. This research, conducted in collaboration with Sinch, evaluates the performance of five state-of-the-art multimodal LLMs in image captioning using standard metrics and human evaluation. Additionally, it explores the emphasis of each model's captions. Do they lean more towards object or action descriptions? Do they include descriptions of the atmosphere in the image and how much? The study shows that BLIP-2 achieves the highest performance based on metrics while Gemini is rated highest in the human evaluation. We also outline limitations in current metrics, where they struggle to assess longer, more detailed captions fairly. We found that GIT and BLIP-2 focus more on objects in an image while Gemini and Qwen2-VL excel at atmospheric descriptions, providing more detailed captions.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/250130_14Tiet.pdf

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

Presenters: Otto Grafström, Ellen Andreasson
Title: Handling Missing Waste Bin Sensor Data Through Machine Learning Predictions
Examiner: Jacek Malec
Supervisor: Pierre Nugues (LTH)

Efficient waste management is a critical challenge for modern societies, with significant environmental and economic implications. The use of sensors to monitor waste container fill levels has proven to be a valuable tool in addressing these challenges. This study focuses on replacing lost sensor data by comparing four predictive modeling approaches: a baseline model, the classical statistical method sARIMAX, a neural network based long short-term memory (LSTM) model, and Chronos-Bolt, a state-of-the-art pre-trained foundation model. The task is divided into two parts: predicting container emptying dates and predicting fill levels at those dates. Among the models, sARIMAX shows the best overall performance, closely followed by Chronos-Bolt. LSTM and the baseline model exhibit significantly lower performance, with LSTM slightly outperforming the baseline. This analysis highlights the continued relevance of classical methods alongside emerging machine learning techniques for reconstructing missing data and supporting efficient waste management practices.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/250130_15AndreassonGrafstrom.pdf