mmdetection/configs/yolox

YOLOX

YOLOX: Exceeding YOLO Series in 2021

Abstract

In this report, we present some experienced improvements to YOLO series, forming a new high-performance detector -- YOLOX. We switch the YOLO detector to an anchor-free manner and conduct other advanced detection techniques, i.e., a decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models: For YOLO-Nano with only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassing NanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors in industry, we boost it to 47.3% AP on COCO, outperforming the current best practice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters as YOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS on Tesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021) using a single YOLOX-L model. We hope this report can provide useful experience for developers and researchers in practical scenes, and we also provide deploy versions with ONNX, TensorRT, NCNN, and Openvino supported.

Results and Models

Backbone	size	Mem (GB)	box AP	Config	Download
YOLOX-tiny	416	3.5	32.0	config	model | log
YOLOX-s	640	7.6	40.5	config	model | log
YOLOX-l	640	19.9	49.4	config	model | log
YOLOX-x	640	28.1	50.9	config	model | log

Note:

1. The test score threshold is 0.001, and the box AP indicates the best AP.
2. Due to the need for pre-training weights, we cannot reproduce the performance of the yolox-nano model. Please refer to https://github.com/Megvii-BaseDetection/YOLOX/issues/674 for more information.
3. We also trained the model by the official release of YOLOX based on Megvii-BaseDetection/YOLOX#735 with commit ID 38c633. We found that the best AP of YOLOX-tiny, YOLOX-s, YOLOX-l, and YOLOX-x is 31.8, 40.3, 49.2, and 50.9, respectively. The performance is consistent with that of our re-implementation (see Table above) but still has a gap (0.3~0.8 AP) in comparison with the reported performance in their README.

Citation

@article{yolox2021,
  title={{YOLOX}: Exceeding YOLO Series in 2021},
  author={Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
  journal={arXiv preprint arXiv:2107.08430},
  year={2021}
}