Abstract
Existing detection-based trackers exploit temporal contexts by updating appearance models or modeling target motion. However, the sequential one-shot integration of temporal priors risks amplifying error accumulation, as frame-level template matching restricts comprehensive spatiotemporal analysis. To address this, we propose SCT-Diff, a video-level framework that holistically estimates target trajectories. Specifically, SCT-Diff processes video clips globally via a diffusion model to incorporate bidirectional spatiotemporal awareness, where reverse diffusion steps progressively refine noisy trajectory proposals into optimal predictions. Crucially, SCT-Diff enables iterative correction of historical trajectory hypotheses by observing future contexts within a sliding time window. This closed-loop feedback from future frames preserves temporal consistency and breaks the error propagation chain under complex appearance variations. For joint modeling of appearance and motion dynamics, we formulate trajectories as unified discrete token sequences. The designed Mamba-based expert decoder bridges visual features with language-formulated trajectories, enabling lightweight yet coherent sequence modeling. Extensive experiments demonstrate SCT-Diff's superior efficiency and performance, achieving 75.4% AO on GOT-10k while maintaining real-time computational efficiency.