Source-linked AI summary
YOLOX: Exceeding YOLO Series in 2021
Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, Jian Sun
TL;DR
YOLO families had not integrated recent advances in anchor-free detection and label assignment. YOLOX combines these techniques with a decoupled head, achieving a better speed–accuracy trade-off across model sizes and 47.3% AP on COCO for YOLOv3, 3.0% above current best practice.
Problem
Recent advances in anchor-free detection and label assignment had not been integrated into YOLO families.
Method
YOLOX combines an anchor-free design, decoupled head, and advanced label assignment strategy to update the YOLO series.
Results
47.3% AP on COCO, surpassing the current best practice of YOLOv3 by 3.0% AP.
Takeaways & Limitations
YOLOX provides a better speed–accuracy trade-off than other counterparts across model sizes and supports practical deployment.
Takeaways & Limitations
The report did not explore larger high-performance YOLO models or important Transformer-based detector features because of time and resource limitations.
Abstract
from arXiv · showhide
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. Source code is at https://github.com/Megvii-BaseDetection/YOLOX.
1. Introduction
YOLOX addresses the gap between recent object-detection advances and YOLO’s anchor-based pipeline by applying these techniques to a practical YOLOv3 starting point. The resulting models improve COCO performance, including 47.3% AP for YOLOX-DarkNet53 and 50.0% AP for YOLOX-L.
- Motivation: Recent advances in anchor-free detection, label assignment, and end-to-end detection had not been integrated into YOLO families.YOLOv4 and YOLOv5 remained anchor-based detectors using hand-crafted training assignment rules.
- Motivation: YOLOX chooses YOLOv3 as its starting point because YOLOv4 and YOLOv5 may be over-optimized for anchor-based pipelines.YOLOv3 remains widely used in industry because of limited computation resources and insufficient software support in practical applications.
- Results: 47.3% AP is achieved by YOLOX-DarkNet53 on COCO at 640 × 640 resolution, surpassing the ultralytics YOLOv3 version’s 44.3% AP.The report describes this as a large-margin improvement over the current best practice of YOLOv3.
- Deployment: YOLOX code is released with ONNX, TensorRT, NCNN, and Openvino support, and a single YOLOX-L model won 1st Place in the Streaming Perception Challenge.The challenge was held at the Workshop on Autonomous Driving at CVPR 2021.
2. YOLOX
YOLOX uses YOLOv3 with Darknet53 as its baseline and improves the detector through a decoupled head, strong augmentation, an anchor-free design, and SimOTA label assignment. These changes yield faster convergence, simpler detection, and substantial COCO AP gains across model variants.
- Baseline: YOLOX adopts YOLOv3 with Darknet53 as its baseline architecture.The baseline includes an SPP layer and several training-strategy changes, including EMA updates, cosine learning-rate scheduling, IoU loss, and an IoU-aware branch.
- Decoupled head: The decoupled head separates classification and regression, converges much faster, and reduces the end-to-end AP decrease from 4.2% to 0.8%.It uses parallel classification and regression branches, with an IoU branch on the regression side.
- Data augmentation: Mosaic and MixUp strengthen augmentation, after which ImageNet pre-training is no more beneficial and subsequent models are trained from scratch.For YOLOX-Nano, removing MixUp and weakening Mosaic improves AP from 24.0% to 25.3%.
- Anchor-free: The anchor-free design reduces predictions per location from 3 to 1, directly predicts box offsets and dimensions, and achieves 42.9% AP.Center sampling assigns the center location of each object to an FPN level using a predefined scale range, reducing detector parameters and GFLOPs.
- SimOTA: SimOTA selects a dynamic number of low-cost predictions within a center region and raises AP from 45.0% to 47.3%, exceeding ultralytics-YOLOv3 by 3.0% AP.It reduces training time and avoids additional Sinkhorn-Knopp solver hyperparameters.
- Model variants: Across YOLOv5-sized models, YOLOX improves AP by ∼3.0% to ∼1.0% with only marginal time increasing, and achieves consistent gains across other backbones.The comparison uses YOLOv5’s modified CSPNet, SiLU activation, PAN head, and scaling rule.
3. Comparison with the SOTA
The comparison with SOTA emphasizes that reported inference speeds are often uncontrolled, while acknowledging larger YOLO models and Transformer-based detectors that reach approximately 60 AP but were not explored due to time and resource limits.
- Comparison methodology: Inference speed varies with software and hardware, so the authors compare YOLO models using the same hardware and code base in a controlled speed/accuracy curve.This addresses the limitations of conventional SOTA comparison tables.
- Comparison with SOTA: ∼60 AP is achieved by current Transformer-based detectors, while larger YOLO models such as Scale-YOLOv4 and YOLOv5-P6 represent high-performance alternatives.The report did not explore these features because of time and resource limitations, though they remain within scope.
4. 1st Place on Streaming Perception Challenge (WAD at CVPR 2021)
YOLOX won first place in the Streaming Perception Challenge at WAD 2021 by optimizing streaming accuracy, which jointly evaluates perception-stack accuracy and latency. The team selected a TensorRT-powered YOLOX-L model meeting the ≤33ms inference-time trade-off for 30 FPS streams.
- Streaming accuracy: Streaming accuracy jointly evaluates accuracy and latency for the entire perception stack at every time instant.The metric accounts for streaming data that should be ignored while computation is occurring.
- Model selection: ≤33ms inference time was identified as the best trade-off point for streaming accuracy on 30 FPS data.This trade-off motivated selecting a powerful model for the challenge.
5. Conclusion
YOLOX is a high-performance anchor-free detector created through updates to the YOLO series. Its decoupled head, anchor-free design, and advanced label assignment improve the speed–accuracy trade-off across model sizes.
- Conclusion: YOLOX forms a high-performance anchor-free detector from updates to the YOLO series.The report identifies YOLOX as the resulting architecture.
- Conclusion: Decoupled head, anchor-free design, and advanced label assignment are the key detection techniques equipped in YOLOX.These techniques are presented as recent advances incorporated into the detector.
- Conclusion: YOLOX achieves a better speed–accuracy trade-off than other counterparts across all model sizes.The passage frames this trade-off as a central outcome of the architecture and its techniques.
- Conclusion: The report highlights boosting YOLOv3, still one of the most widely used detectors in industry.The supplied passage begins describing this result but is truncated before giving the quantitative outcome.