In this thesis project you will be working with Veoneer in Linköping, in collaboration with the division for Media and Information Technology at Campus Norrköping.
Modern deep learning-based traffic sign recognition algorithms rely on large amount of training data and large amounts of validation/test data. Traffic sign recognition with support for many sign classes is an inherently unbalanced problem, where some sign classes are extremely rare in the real world and might only exist in a few countries. This means that collecting large amounts of data for rare classes is difficult and expensive. One possible way to alleviate this problem could be to insert synthetic sign patches into real images and use GANs to bridge the domain gap between the augmented samples and the real samples.
The goal of this master thesis is to investigate different approaches to augment images using synthetic sign patches for use in TSR CNN trainings to increase performance, in particular for uncommon sign classes. The suggested approach is to use CycleGANs, possibly with some extra tricks to ensure that signs are not removed and not distorted in a way that changes the sign class. Other simpler methods of blending inserted sign patches can be used as baselines.
Students with a background in computer science, electrical engineering applied mathematics or similar, with experience from courses in computer vision, machine learning and in particular deep learning approaches to computer vision tasks
Contact: Gabriel Eilertsen (firstname.lastname@example.org), division of Media and Information Technology, Department of Science and Technology, Campus Norrköping, Linköping University