Center3dAugNet: Effect of Rotation Representation on One-Stage Joint Car Detection and 6D-Pose Estimation

Abstract

Joint car detection and 6D-pose estimation on monocular images is a topic of active research. This paper discusses the fast one-stage method based on CenterNet deep architecture, named Center3dAugNet. Particular attention is paid to the study of rotation representation effect on network training results. It is of high importance due to discontinuity of 3D-object yaw, pitch and roll angles. Relative quaternion representation has shown the greatest positive effect on the mean average precision of network. Also we introduce new loss function for rotation estimation in terms of quaternions. The performance of the proposed approach is evaluated both on powerful hardware platforms based on NVidia Tesla V-100, RTX2080Ti, RTX2070, as well as on embedded computing boards Jetson Xavier and TX2. Our experiments indicate that the presented approach is suitable for real-time computer vision systems of unmanned vehicles and mobile ground robots. Source code is made publicly available https://github.com/cds-mipt/center_3d_aug_net.