Horse ReIDing: Addressing Re-Identification in Horse Racing Scenarios

Re-Identification (ReID) tasks, traditionally employed in person tracking across diverse camera views, face unique challenges in the domain of horse racing due to frequent occlusions, dynamic motion, and varying environmental conditions. This study addresses these complexities by developing a custom pipeline and dataset for jockey ReID, specifically collected from horse racing footage. A ResNeXt-based architecture is employed to process input data, with additional experiments exploring the inclusion of segmentation mask information for improved performance. Empirical evaluations demonstrate the model's efficacy in both closed-set and open-set scenarios, showcasing significant gains in mean Average Precision (mAP) and top-k Cumulative Matching Characteristic (CMC) rank metrics when segmentation masks are incorporated. Comparative analysis across different ResNeXt configurations underscores the robustness and scalability of the proposed approach, contributing as a pioneering framework for ReID in high-motion sports contexts and advancing the application of computer vision technologies in horse racing scenarios.