LAUSR

3D single object tracking in point clouds is an essential task in robotics and autonomous driving. Many astonished trackers only adopt the voting-based region proposal network (RPN) to regress the object's location. However, they suffer from heavy outlier votes and ignore the global semantic features of targets. To resolve the problems, we propose a two-stage RPN module with local-to-global feature refinement for accurate tracking in point clouds. Specifically, deep Hough voting is applied to obtain coarse proposals in the first stage. In the second stage, we design a local feature refinement (LFR) module and a global feature refinement (GFR) module to realize accurate localization jointly. The LFR module excludes noisy outliers in disordered point clouds and obtains refined local features for coarse proposals. After that, the GFR module explores the relationships among all proposals to weigh the proposal-wise global context features. Integrating the proposed two-stage RPN module into the previous method BAT Zheng et al. (2021), we develop a coarse-to-fine 3D single object tracker in point clouds abbreviated as C2FT. Extensive experiments on KITTI and nuScene benchmarks demonstrate that C2FT achieves favorable performance with a real-time speed ($\sim$50 FPS). Furthermore, the proposed LFR and GFR modules are generalized and can be easily integrated into other trackers.