YOLOX: Exceeding YOLO Series in 2021

  • 2021-07-18 12:55:11
  • Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, Jian Sun
  • 89


In this report, we present some experienced improvements to YOLO series,forming a new high-performance detector -- YOLOX. We switch the YOLO detectorto an anchor-free manner and conduct other advanced detection techniques, i.e.,a decoupled head and the leading label assignment strategy SimOTA to achievestate-of-the-art results across a large scale range of models: For YOLO-Nanowith only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassingNanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors inindustry, we boost it to 47.3% AP on COCO, outperforming the current bestpractice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters asYOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS onTesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place onStreaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021)using a single YOLOX-L model. We hope this report can provide useful experiencefor developers and researchers in practical scenes, and we also provide deployversions with ONNX, TensorRT, NCNN, and Openvino supported. Source code is athttps://github.com/Megvii-BaseDetection/YOLOX.


Quick Read (beta)

loading the full paper ...