0%

YOLOX

发表于 更新于 分类于 阅读次数： Valine：

Abstract

  • 引入anchor-free
  • decoupled head ,simOTA
  • Performance: yolo-nano: 1.8% ap； yolov3 : 3%; YOLOv5-L:1.8% ;
  • provide tensorrt onnx version

Introduction

  • 近两年目标检测的研究热点： anchor-free； advanced label assignment strategies[37,36,12,41,22,4]; NMS-free detectors[2,32,39]
  • YOLOv3 as default
  • we boost the YOLOv3 to 47.3%AP (YOLOX-DarkNet53) on COCO with 640 × 640 resolution, surpassing the current best practice of YOLOv3(44.3% AP, ultralytics version2) by a large margin.’
  • YOLOv5 640x640； 50.0% AP, supass 1.8%AP

YOLOX

Implementation details

  • 300 epoch , 5 epoch warmup
  • lr = init_lr * (batchsize/64), init_lr = 0.01, cos lr schedule, (8-gpu, batchsize:128)
  • input size : 448 to 832 , 32 strides

YOLOv3 baseline

  • DarkNet53 + SPPlayer
  • ==adding EMA weights updating, cosine lr schedule, IoU loss and IoU-aware branch. We use BCE Loss for training cls and obj branch, and IoU Loss for training reg branch.==

Decoupled head

image-20210918121226730

image-20210918121247956

Strong data augmentzation

  • Mosaic and Mixup, and closed it for the last 15 epoches
  • After using strong data augmentation, ImageNet pre-trained is no more beneficial , == train all the following models from scratch.==

Anchor-free

Multi positives

  • not only the center point as positive ; center 3x3 area as positives ; as ‘center sampling’ in Fcos, balance the positive/negative samplings.

SimOTA：

  • ==Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, 2021.==
  • simOTA是在上面基础上的优化，去掉了Sinkhorn-Knopp 算法，选择topk替代。
  • OTA的主要作用是，对训练的样本的动态分配，可以在训练的时候，自动计算正样本；可参考博客：https://zhuanlan.zhihu.com/p/394392992

aug的提升效果：

image-20210922111546338

Other backbone

vs yolov5

  • use cspnet , silu activation , pan head

image-20210922114054103

vs tiny, nano detector

image-20210922114234135

model size and data augmentation

  • 对于小模型，需要减少augment(mosaic,mixup等)；对于大模型需要增强augment

总结

  • 本文依然是在yolov3的基础上进行一系列的魔改，增加了anchor free；decouple head; augments等，最终提升了模型的效果，在大模型和小模型中，都取得了一定的优势。