Henrique’s research project focused on studying bird’s eye view (BEV) perception using only dashcam footage. BEV semantic maps provide a representation of the environment with several desirable properties for assisted and automated driving and robotics. He addressed the challenge of limited or unavailable BEV-labeled data motivated by a real industrial application and the company where his project was carried out (Verizon Connect). He developed RendBEV, the first self-supervised training methodology that allows BEV segmentation models to be supervised without any direct BEV supervision.
During his PhD, he also collaborated in projects to collect a dashcam dataset to study the impact of viewpoint shifts, generating a synthetic dataset with the same scope, and identify manoeuvres in compressed representations of video known as motion profiles.
