TAPTrack: Spatiotemporal Tracking With Target-Aware Prompt Learning

Effective spatiotemporal prompts are crucial for robust single object tracking. However, existing approaches often model the appearance and temporal information of the template and search region jointly, which not only introduces additional computational overhead but also increases background interference, thereby hindering precise target perception. To address these issues, we propose a simple yet effective spatiotemporal tracker, named TAPTrack, which enhances the understanding of the target solely through search features and further aggregates temporal information to capture the target's historical context. Specifically, a set of learnable tokens is introduced to extract temporal features of the target. We then incorporate an incomplete attention mechanism into the Transformer backbone to decouple the modeling of temporal tokens, templates and search regions, while preserving rich interactions within the search branch. After backbone feature extraction, a target aggregation module is employed to further fuse the temporal and template features, thereby reinforcing target-aware temporal cues. These refined cues are then used to modulate the search features and guide subsequent network propagation for accurate target localization. Extensive experiments demonstrate that TAPTracker achieves competitive performance on multiple challenging tracking benchmarks while running at a real-time speed of 115 FPS. The source code and models will be released on https://github.com/xiaomengxin123/TAPTrack.