Choosing a detector for multi-pedestrian tracking involves a clear trade-off between accuracy and compute, according to a new study that compares three detector families across three crowded-scene benchmarks. The authors evaluated YOLO11, RT-DETR, and YOLO26 detectors paired with the ByteTrack and BoT-SORT trackers on the MOT17, MOT20, and SOMPT22 datasets. RT-DETR consistently achieved the highest mean HOTA scores—a composite tracking quality metric—with both trackers, and its advantage was most pronounced on the densely crowded MOT20 set. That accuracy came at a price: RT-DETR required noticeably more computation.
For applications where speed matters, the results point to YOLO26 variants, particularly YOLO26-L with standard non-maximum suppression (NMS), as offering a strong balance of quality and efficiency on MOT17 and SOMPT22. The study also tested whether YOLO26's end-to-end inference mode—which removes the need for NMS—could improve tracking. It did not: end-to-end inference produced no consistent gains over the NMS-based regime. The authors frame the findings as a practical guide, suggesting that detector choice should depend on whether the priority is peak accuracy or real-time throughput.