Answers are grounded in this paper and cite their sources.
Diffusion-based detection models formulate object detection as an iterative denoising process from noisy boxes to target bounding boxes, offering a promising paradigm for sports video object detection. However, two critical challenges remain. (1) Existing diffusion detectors perform frame-independent denoising from random noise, disregarding the temporal continuity inherent in video sequences, which causes detection results for the same target to exhibit spatial jitter across consecutive frames. (2) In densely crowded regions common in sports scenarios, the self-attention mechanism within the denoising decoder causes spatially adjacent proposal representations to progressively converge during iterative denoising, ultimately collapsing onto the same target location and resulting in missed detections. To address these issues, we propose TDDiffDet, a framework comprising two core modules: the Temporal Prior-guided Denoising module (TPDen) and the Density-Aware Decoupling module (DADec). TPDen leverages detection results from the previous frame as denoising initialization priors and injects inter-frame motion features into the conditioning pathway of the denoising network, enabling refinement from meaningful priors rather than pure noise, thereby improving temporal consistency while reducing redundant computation. DADec introduces a density-aware coupling suppression mechanism into the self-attention layers of the denoising decoder, dynamically attenuating attention associations between adjacent proposals in densely crowded regions to preserve the independence of each proposal’s denoising trajectory and prevent dense targets from being merged due to representation coupling. Extensive experiments on SportsMOT and SoccerNet-Tracking demonstrate that TDDiffDet achieves significant performance improvements over existing methods, validating the effectiveness of the proposed strategies.
No takes yet. Share an insight, caveat, or question.
Qi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: