Basketball Player Re-Identification Based on Appearance and Number Distance Fusion with Dynamic Gallery Update

Player re-identification (Re-ID) is a core technique for turning basketball broadcast

video into player-level statistics, yet it faces two obstacles: teammates wear very sim- ilar jerseys, which degrades appearance-only discrimination between them, and high- accuracy Re-ID normally requires large, expensively annotated galleries. This thesis

presents a training-free query–gallery Re-ID pipeline that identifies players directly from bounding boxes extracted across full-length broadcasts. Appearance features from a CLIP encoder are fused with jersey-number features— localized by a fine-tuned YOLOv8 detector and embedded by a ResNet-50—through a median-absolute-deviation (MAD) alignment that makes the two heterogeneous distances directly comparable and gracefully falls back to appearance when no number is legible. A label-free cross-team penalty, driven by estimated team affiliations, suppresses confusions between same-number players on opposing teams. Finally, a confidence-gated dynamic gallery grows online from a minimal seed of five images per identity: a consensus gate admits only high-confidence queries, and a dispersion-based eviction keeps the retained samples diverse. On a 22-game Basketball Re-ID dataset spanning NBA, Olympic and 3x3 footage, the method attains a top-1 accuracy of 94.90%, outperforming TransReID, CLIP-ReID, BPB and PRT while requiring no training on the target footage.