Presented our findings on online and offline real-time tracking at MILA

In this study, we investigate the performance of the YOLOv8 single-stage object detection model in Multiple Object Tracking (MOT) within ice hockey. We integrate YOLOv8 both with on- line (SORT, DeepSORT) and offline, graph-based (MOT Neural Solver) tracking algorithms, evalu- ating on the MOT17 and McGill Hockey Track- ing Dataset (MHTD) 2020. Our key finding is that the finetuned YOLOv8x model, combined with the Neural Solver, achieved above MOT17 benchmark tracking performance on MHTD with a HOTA score of 0.776, showcasing its suitability for dynamic and occlusion-prone environments. Additionally, YOLOv8m demonstrated effective tracking with greater efficiency. These results highlight the potential of single-stage models in specialized MOT applications. Click here for the full report