A Lightweight Deep Learning Model for Vehicle Viewpoint Estimation from Dashcam Images

Abstract

Vehicle viewpoint estimation from vehicle cameras is a crucial component of road scene understanding.In this paper, we propose a deep lightweight method to predict vehicle viewpoint from a single RGB dashcam image. To this aim, we customize and adapt state-of-the-art deep learning techniques for general object viewpoint estimation to the vehicle viewpoint estimation task. Furthermore, we define a novel objective function that takes into account errors at different granularity to improve neural network training. To keep the model lightweight and fast, we rely upon MobileNetV2 as backbone.Tested both on benchmark viewpoint estimation data (Pascal3D+) and on actual vehicle camera data (nuScenes), our method is shown to outperform the state of the art in vehicle viewpoint estimation, in terms of both accuracy and memory footprint.