DenShift‐Det: Density‐Shift Adaptive Feature Encoding and Individual Boundary Discrimination for Sports Scene Object Detection

ABSTRACT

Object detection in sports videos serves as a fundamental task for high‐efficiency sports video analysis, tactical analysis, and athlete tracking. However, existing methods face two critical challenges arising from the drastic intra‐game density variations and highly homogeneous player appearances inherent in sports scenarios. (1) At the feature encoding level, sparse and dense scenes impose contradictory requirements on feature representation: isolated targets in sparse regions require large receptive fields to gather surrounding scene context for reliable identification and false positive suppression, whereas tightly packed targets in dense regions demand fine‐grained local features to preserve individual boundaries and prevent detection merging; current encoders apply uniform feature extraction regardless of local density, systematically failing at both extremes. (2) At the instance discrimination level, players from the same team often wear identical uniforms and exhibit similar body shapes, causing adjacent targets to have highly similar visual features. Existing detectors mainly rely on appearance cues and therefore tend to merge neighbouring homogeneous players into a single response under occlusion or contact. To address these issues, we propose DenShift‐Det, a framework comprising two complementary modules: the Density‐Adaptive Feature Encoding module (DAFE) and the Individual Boundary Discrimination Module (IBDM). Specifically, DAFE estimates a fine‐grained local density map from multi‐scale encoder features and uses density‐conditioned multi‐granularity feature encoding to activate broad contextual aggregation in sparse regions while preserving compact boundary‐sensitive features in dense regions. IBDM extracts pose‐aware instance descriptors and spatial relation encodings from candidate targets, performs graph contrastive learning over neighbouring target pairs, and injects the resulting discriminative features into decoder cross‐attention so that each query can actively distinguish itself from nearby homogeneous targets. Experiments on SportsMOT and SoccerNet‐Tracking demonstrate that DenShift‐Det achieves significant performance improvements over existing methods, with particularly notable gains in density‐extreme subsets, validating the effectiveness of the proposed density‐adaptive and boundary‐discriminative design.