HyGFNet: Hybrid Geometry-Flow Learning Network for 3D Single Object Tracking

3D single object tracking (SOT) which attempts to accurately locate the target object in the current frame, has made significant advancements over past years. However, most previous works built upon the Siamese architecture usually focus on the learning and matching of geometry information, while neglecting the motion information of the target. Consequently, those methods face challenges when distinguishing the target object from similar distractors. To overcome this limitation, we enhance the architecture by incorporating flow estimation, presenting a novel multi-frame hybrid geometry-flow learning network for 3D SOT. The proposed framework exploits both geometry and flow information from historical frames and further integrates the learned information into the current frame to improve the target object localization. Specifically, we propose a geometry matching branch and a flow estimation branch. The geometry matching branch first captures the geometry feature of each frame and then aligns these features through multi-frame spatial-temporal matching, leading to a geometry-aware feature for the target object. Meanwhile, in the flow estimation branch, a multi-frame flow-aware enhancement module is proposed to explicitly capture flow information across frames, leading to a flow-aware feature for the target object. Finally, the geometry-aware and flow-aware features are fused with the original feature to predict the position of the target object. Extensive experiments conducted on the KITTI and nuScenes datasets validate the effectiveness and show the competitive performance of our method. The code will be open soon.