YOLOX: Exceeding YOLO Series in 2021

The report presents YOLOX, a set of experienced improvements to the YOLO series of object detectors. It switches the YOLO detector to an anchor-free manner and incorporates advanced detection techniques, notably a decoupled head and the leading label assignment strategy SimOTA. These design choices are applied consistently across a wide range of model scales.

YOLOX achieves state-of-the-art trade-offs: YOLO-Nano with only 0.91M parameters reaches 25.3% AP on COCO (surpassing NanoDet by 1.8%), YOLOv3 is boosted to 47.3% AP, and YOLOX-L reaches 50.0% AP at 68.9 FPS on a Tesla V100, exceeding YOLOv5-L by 1.8% AP. A single YOLOX-L model won first place in the Streaming Perception Challenge at the CVPR 2021 Workshop on Autonomous Driving, and the authors provide deployable versions supporting ONNX, TensorRT, NCNN, and OpenVINO for practical use.