Mono6D++: Learning Point Cloud Visibility for 3D Prior-based Vehicle 6D Pose Estimation

Yangxintong Lyu, Olivier Ducastel, Remco Royen, Adrian Munteanu

Abstract: Point cloud visibility is a crucial attribute for 3D tasks as it links the visible object points to a given viewpoint. In this paper, we address the problem of point cloud visibility for monocular vehicle 6D pose estimation. To this end, a network, dubbed Mono6D++, is introduced which jointly predicts vehicle poses and the associated points visibility. Our method mainly consists of: 1) a multi-model feature extraction module and 2) a fusion unit for learning the pose- and visibility-specific representations. Consequently, the proposed method significantly outperforms the baseline approaches. Mono6D++ is capable of handling heavily occluded, truncated and/or appearance-ambiguous vehicles.