Towards End-to-End Deep Learning for Visual Detection and Tracking

Visual object detection and tracking is a complex computer vision problem that often requires multiple components and can be decomposed in various ways. Traditional deep learning methods divide the high-level problem into separate components optimized for surrogate tasks and combine them heuristically during inference. In this thesis, we explore a different methodology that decomposes the problem in a way that as many aspects of the problem can be addressed by an integrated machine learning process. As a result, the model is learnt to directly optimize the high-level problem end-to-end. This enables knowledge sharing between sub-task modules and streamlines the model.