Master Thesis - Damage detection in used clothes - Under tillsättning eller tillsatt, ej sökbar

Project goal
Build an object detection/segmentation model using SOTA models like the Segment Anything Model (SAM) or variants thereof to detect damage in used clothes.

Background and purpose
The fashion industry is shifting its focus from solely creating new clothing to promoting the reuse and recycling of existing garments. Our research encompasses two projects, one backed by Vinnova and the other by the EU, aimed at harnessing the power of AI for automating clothing reuse. Traditional human sorters typically undergo extensive training lasting several months to assess the best way to repurpose clothing items, and their decisions are often prone to errors. We are investigating how AI can enhance the sorting process by minimizing errors and expediting decision-making, ultimately optimizing the reuse of clothes.

About the project
The project will commence by establishing a comprehensive understanding of the current state-of-the-art methods in object detection and segmentation, which may entail reviewing and implementing research papers showcasing cutting-edge solutions. Notably, some publicly available large models like SAM exhibit respectable zero-shot performance but may lack finer-grained details, necessitating additional annotation efforts. Time permitting, there is also a plan to conduct a preliminary investigation into damage detection using multi-spectral cameras.

Our primary focus will be on employing a clothing dataset created in collaboration with Wargön Innovation AB, as part of our ongoing projects. In collaboration with researchers from RISE Linköping, the student will undertake the training of machine learning models on fashion data and assess their performance using various metrics. It is worth noting that we provide access to GPU resources for model training and evaluation. However, it's important to mention that our GPU resources primarily consist of consumer-grade GPUs.

This project centers on the crucial task of damage detection within the realm of clothing, primarily through the utilization of object detection and segmentation techniques.

Requirements/knowledge
The student is expected to have some background in deep learning, computer vision and some experience in training models. We use the Pytorch framework.

Terms
Credits: 30 ECTS (in agreement with the examiner)
Location: Linköping
Start: Jan. 2024

The selected candidate will receive a fixed payment of 30,000 SEK upon the successful completion of the project.

Contact
Farrukh Nauman, PhD, Applied AI and IoT, RISE Linköping (farrukh.nauman@ri.se)
Per Bröms, PhD, Applied AI and IoT, RISE Linköping (per.broms@ri.se)

Applications will be evaluated continuously, and the start date will be agreed with the successful applicant(s). We especially encourage underrepresented groups to apply and contribute their diverse perspectives to this valuable project.

Last application date is 22 October, 2023.