Real-Time 3D Single Object Tracking With Transformer

LiDAR-based 3D single object tracking is a challenging issue in robotics and autonomous driving. Currently, existing approaches usually suffer from the problem that objects at long distance often have very sparse or partially-occluded point clouds, which makes features extracted by model ambiguous. Ambiguous will make it hard to locate target finally lead bad results. To solve this problem, we utilize powerful Transformer architecture propose Point-Track-Transformer (PTT) module for cloud-based task. Specifically, PTT generates fine-tuned attention computing weights, guides tracker focusing on important of improves ability complex scenarios. evaluate our module, embed into dominant method construct novel SOT named PTT-Net. In PTT-Net, voting stage proposal generation stage, respectively. could interactions among patches, learns context-dependent features. Meanwhile, capture contextual information between background. We PTT-Net KITTI NuScenes datasets. Experimental results demonstrate effectiveness superiority surpasses baseline noticeable margin, $\sim$ 10% Car category. also has significant performance improvement general, combination transformer pipeline enables achieve state-of-the-art both two Additionally, run real-time 40FPS NVIDIA 1080Ti GPU. Our code open-sourced research community https://github.com/shanjiayao/PTT .