A Lightweight and Detector-Free 3D Single Object Tracker on Point Clouds

Recent works on 3D single object tracking treat the task as a target-specific detection task, where an off-the-shelf detector is commonly employed for tracking. However, it non-trivial to perform accurate since point cloud of objects in raw LiDAR scans usually sparse and incomplete. In this paper, we address issue by explicitly leveraging temporal motion cues propose DMT, Detector-free Motion-prediction-based Tracking network that completely removes usage complicated detectors lighter, faster, more than previous trackers. Specifically, prediction module first introduced estimate potential target center current frame point-cloud-free manner. Then, explicit voting proposed directly regress box from estimated center. Extensive experiments KITTI NuScenes datasets demonstrate our DMT can still achieve better performance ( $\sim $ 10% improvement over dataset) faster speed (i.e., 72 FPS) state-of-the-art approaches without applying any detectors. Our code released at https://github.com/jimmy-dq/DMT .