JEDE: Universal Jersey Number Detector for Sports

Contrastive Learning for Sports Video: Unsupervised Player Classification

“…The problem of jersey number recognition has been posed as image-level recognition [6,17,19,20,25] as well as tracklet-level recognition [4,7,26,27]. Some methods detect and localize the jersey number region and then classify the numbers [17,19,20], while others assume that the im-age region containing the jersey number has already been cropped [6,12,25]. Progress on this problem has been slowed by the lack of large public datasets that can be used to compare methods.…”

“…Gerke et al [12] and Li et al [17] were among the first to apply CNN-based classification approaches to image-level jersey number recognition, and CNNs have been the dominant approach since this time. Liu et al [19,20] demonstrated the utility of body pose detection to improve classification with Faster R-CNN [22] and Mask R-CNN [16] architectures, respectively.…”

“…In contrast to previously proposed jersey number STR approaches [8,21] our method does not require jersey number bounding box annotations. As in [19,20], we take advantage of body pose detection to localize the jersey number. As in [25] we use a weak-labelling strategy to generalize from image-level to tracklet-level annotation.…”