Weakly-supervised Computer Vision Methods for Autonomous Driving

Abstract

The thesis summarises the author’s post-PhD scientific work in the field of computer vision for autonomous driving. The common theme across the presented methods is the reduction of the reliance on costly human annotation, which is one of the main obstacles to training reliable perception models at scale. Instead of labelled data, the methods draw supervision from signals that are essentially free to obtain – the laws of physics, projective geometry, the temporal consistency of a scene observed over time, or the predictions of simpler, readily available off-the-shelf models. Across several autonomous driving tasks, models trained without domain-specific human labels are shown to match or closely approach the accuracy of their fully-supervised counterparts, while being substantially cheaper to supervise.