DPTrack: Dual-prompt guided network for visual object tracking

Recently, prompt-based spatio-temporal trackers have achieved impressive advancements. However, these methods tend to incorporate only limited spatio-temporal information through prompts, such as recursive prompts or historical prompts, which fail to fully exploit spatio-temporal context information in video sequences, i.e., the instantaneous states of nearby frame and the consistent states of past frames. This oversight inhibits further performance improvement in complex scenes. To tackle this issue, we resort to prompt learning and present a dual-prompt guided visual tracking network (DPTrack), consisting of an instantaneous prompt network (IPN) and a spatio-temporal consistency prompt network (ST-CPN). Specifically, the IPN captures the appearance changes of the nearby frame to generate instantaneous prompt, which are directly involved in feature extraction and information interaction. The ST-CPN learns a set of learnable prompts to summarize the spatio-temporal consistency of previous frames, and then iteratively guides the search features to emphasize the target embedding. In this way, the dual-prompt exploits rich spatio-temporal cues, enhancing the adaptability and robustness of the tracker. Furthermore, we introduce a spatio-temporal pooling collection mechanism (SPC) to maintain consistency and adapt to the appearance changes. Extensive experiments on seven benchmarks prove that the proposed DPTrack achieves very promising tracking performance. Our DPTrack achieves 77.6 % AO on GOT-10k and 52.8 % AUC on LaSOT${}_{ext}$. Notably, it obtains 73.9 % AUC, 81.8 % P, and 84.5 % P${}_{Norm}$ on LaSOT, 60.8 % EAO, 78.2 % A, and 89.3 % R on VOT2020, outperforming the remarkable trackers and demonstrating its superiority.