SAMIDARE: Advanced Tracking-by-Segmentation for Dense Scenarios

Researchers Shozaburo Hirano and Norimichi Ukita have introduced SAMIDARE, a novel framework designed to improve multi-object tracking in dense, crowded scenarios—particularly for automated sports analysis. The method tackles persistent weaknesses in segmentation-based tracking, such as mask errors and frequent ID switches when objects overlap or leave the frame temporarily. SAMIDARE builds on the SAM2MOT baseline with three key innovations: density-aware mask re-generation, selective memory updates, and state-aware association with new track initialization. These components work together to adaptively manage mask features, ensuring target integrity even under heavy occlusion.

Evaluated on the SportsMOT dataset, SAMIDARE achieves state-of-the-art performance, outperforming the baseline by 2.5 HOTA (Higher Order Tracking Accuracy) and 4.2 IDF1 (identification F1 score) points on the validation set. The results demonstrate that adaptive feature management and state-aware association provide a robust, efficient solution for dense sports tracking. The code is publicly available, offering a practical tool for sports analytics and other high-density tracking applications.