YOLOv10: Real-Time End-to-End Object Detection

YOLOv10 advances the YOLO family of real-time object detectors by tackling two limitations: the dependence on non-maximum suppression (NMS) for post-processing, which blocks end-to-end deployment and increases inference latency, and computational redundancy across insufficiently examined model components. The authors introduce consistent dual assignments to enable NMS-free training and apply a holistic efficiency-accuracy driven strategy to redesign the model's components from both efficiency and accuracy perspectives.

Experiments show YOLOv10 achieves state-of-the-art performance and efficiency across model scales. For instance, YOLOv10-S runs 1.8x faster than RT-DETR-R18 at similar COCO AP while using 2.8x fewer parameters and FLOPs, and YOLOv10-B has 46% less latency and 25% fewer parameters than YOLOv9-C for the same performance, pushing the speed-accuracy frontier for deployable detectors.