PBMA: Enhancing 3D point cloud tracking with Point-to-Box Motion Augmentation

3D single object tracking (SOT) holds significance in various applications, especially within the domain of autonomous driving. The 3D SOT task involves the processing of point cloud sequence data that comprises rich inter-frame spatio-temporal information and target motion cues. However, contemporary methods often involve scene cropping and coordinate system transformations, leading to under-utilize of the spatio-temporal context and disrupt the intrinsic motion coherence between successive frames. To tackle this problem, we propose Point-to-Box Motion Augmentation (PBMA), an intuitive method that incorporates coordinate shifts and feature displacement across consecutive frames to augment the modeling of target motion. In particular, our approach estimates inter-frame scene flow and feature flow, which are jointly referred to low-level point motion. This enables a comprehensive understanding of the dynamic environment and the capture of motion cues from the spatio-temporal context, all without the need for cropping or coordinate system transformations. Ultimately, PBMA translates low-level point motions and geometry features for augmenting high-level target motion modeling. Extensive experiments on large-scale datasets show that PBMA achieves state-of-the-art tracking performance while running at 25 FPS.