ViewAdapt-Det: view-adaptive detection for soccer broadcasting videos

Object detection in soccer videos plays an important role in intelligent broadcasting, tactical analysis, and player tracking. Frequent view switching driven by multi-camera coordination in soccer broadcasts introduces two distinct challenges for existing detection methods. First, in the temporal dimension, shot transitions cause abrupt visual changes between adjacent frames. Existing temporal aggregation methods forcibly fuse features from heterogeneous views, injecting cross-view noise that can degrade detection accuracy below single-frame baselines. Second, in the feature dimension, target appearance varies significantly across views: players in panoramic shots are low-resolution small objects characterized mainly by coarse contours and color, whereas the same players in close-up shots become high-resolution large objects with rich texture and fine-grained detail. A unified detection strategy struggles to handle both extremes simultaneously. We propose ViewAdapt-Det, a framework consisting of two complementary modules. The View-Aware Temporal Gating module (VATG) detects shot transitions and dynamically controls temporal aggregation intensity, suppressing historical frame contributions during view changes while preserving temporal enhancement during continuous sequences. The View-Conditioned Detection Modulation module (VCDM) infers the current view type and conditionally modulates detection features, allowing the detector to apply view-specific feature processing strategies that match the visual characteristics of each viewpoint. Experiments on SoccerNet-Tracking, SportsMOT, and our self-constructed BroadcastSwitch-Soccer dataset show that ViewAdapt-Det outperforms existing methods and demonstrates stronger robustness under abrupt shot transitions.