GLT-T: Global-Local Transformer Voting for 3D Single Object Tracking in Point Clouds

Current 3D single object tracking methods are typically based on VoteNet, a region proposal network. Despite the success, using seed point feature as cue for offset learning in VoteNet prevents high-quality proposals from being generated. Moreover, points with different importance treated equally voting process, aggravating this defect. To address these issues, we propose novel global-local transformer scheme to provide more informative cues and guide model pay attention potential points, promoting generation of proposals. Technically, (GLT) module is employed integrate object- patch-aware prior into features effectively form strong representation geometric positions thus providing robust accurate learning. Subsequently, simple yet effective training strategy designed train GLT module. We develop an prediction branch learn treat output weights vector constraint term. By incorporating above components together, exhibit superior method GLT-T. Extensive experiments challenging KITTI NuScenes benchmarks demonstrate that GLT-T achieves state-of-the-art performance task. Besides, further ablation studies show advantages proposed over original VoteNet. Code models will be available at https://github.com/haooozi/GLT-T.