BID-Det: Bias-Aware and instance-discriminative detection for soccer scenarios

Abstract

Football video object detection is a fundamental task underpinning applications such as intelligent broadcasting, tactical analysis, and player tracking. Existing methods face two key challenges in football scenarios. (1) Localization bias in crowded regions: football players frequently gather in dense formations and occlude one another, making it difficult for the regression branch to distinguish the independent boundaries of individual targets. This often leads to shifted bounding-box predictions, while a single regression head lacks explicit modeling of residual patterns in crowded regions. (2) Insufficient instance-level discriminability among players from the same team: players on the same team wear identical uniforms, resulting in highly similar feature representations. Such similarity can mislead the model into confusing candidate responses from adjacent but distinct players, thereby increasing the risk of missed detections. To address these challenges, we propose BID-Det, a detection framework comprising two core modules: the Bias-Aware Dual-Head Correction (BADC) module and the Position-Guided Instance Discrimination (PGID) module. BADC introduces a parameter-independent auxiliary residual head to explicitly learn density-weighted localization residuals between the predictions of the primary head and the matched ground-truth boxes. These residuals are then used as differentiable correction terms for candidate boxes, enabling the detector to compensate for localization bias caused by dense occlusion. PGID imposes instance-discriminative constraints through instance-aware contrastive learning: it pulls together candidate representations matched to the same ground-truth target while pushing apart representations of spatially adjacent candidates that belong to different ground-truth targets of the same class. Moreover, a consistency loss between feature distance and spatial distance is introduced to strengthen separability among different instances, ensuring that neighboring targets of the same class remain distinguishable in the candidate feature space. Experiments on the SoccerNet-Tracking and SportsMOT datasets demonstrate that BID-Det achieves substantial performance gains over the strongest baseline, validating the effectiveness of the proposed method.