Title:	Representation Learning for Computer Vision
DNr:	Berzelius-2025-25
Project Type:	LiU Berzelius
Principal Investigator:	Atsuto Maki <atsuto@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2025-02-01 – 2025-08-01
Classification:	10207
Keywords:

Abstract

Deep learning-based computer vision for manufacturing quality inspection is constrained by the cost and effort of acquiring large amounts of annotated training data. This project aims to overcome these limitations by training models solely on synthetic data generated from manufacturing CAD models. The focus is on enabling models to learn robust object representations that bridge the domain gap between synthetic and real-world data, ensuring reliable performance in real-world applications. In the previous Berzelius project, we utilized domain randomization and active learning strategies to address this challenge. In this project, we aim to expand the scope of representation learning through the following objectives: 1). Extending synthetic active learning to additional manufacturing use cases. Including more complicated manufacturing environments. 2). Advancing research on open vocabulary vision models by exploring efficient self-distillation methods and leveraging vision-language models to enhance representation learning. 3). Incorporating self-supervised learning to further understand representations in image and video data. 4). Expanding the research scope beyond manufacturing to other industries, including football, by exploring representation learning that understands game dynamics and utilizing synthetic data to support it. This project is a collaboration between Scania CV AB and KTH. It is partially supported by the Wallenberg AI, Autonomous Systems, and Software Program (WASP) funded by the Marianne and Marcus Wallenberg Foundation.