A New Approach to Learning of 3D Characteristic Points for Vehicle Pose Estimation

Abstract

This article discusses the challenges of estimating the pose of a vehicle from monocular images in an uncontrolled environment. We propose a new neural network architecture that learns 3D characteristic points of vehicles from image crops and coordinates of 2D keypoints on images. To facilitate supervised training of this network, we pre-process the ApolloCar3D dataset to obtain labelled 3D characteristic points of different car models. We evaluate our approach on the ApolloCar3D benchmark and demonstrate results competitive to state-of-the-art methods.