yolov5自适应anchors

在进行yolov5训练的时候，会输出：

Analyzing anchors... Best Possible Recall (BPR) = 0.8838. Attempting to generate improved anchors, please wait...
WARNING: Extremely small objects found. 2274 of 14719 labels are < 4 pixels in width or height.
Running kmeans for 9 anchors on 14700 points...
thr=0.25: 0.9927 best possible recall, 5.10 anchors past thr
n=9, img_size=480, metric_all=0.348/0.766-mean/best, past_thr=0.515-mean: 10,3,  18,6,  24,12,  36,16,  42,28,  62,39,  77,68,  110,49,  121,100
Evolving anchors with Genetic Algorithm: fitness = 0.7855: 100%|████████████| 1000/1000 [00:02<00:00, 484.50it/s]
thr=0.25: 0.9972 best possible recall, 5.23 anchors past thr
n=9, img_size=480, metric_all=0.358/0.785-mean/best, past_thr=0.524-mean: 11,2,  16,5,  22,8,  28,12,  36,18,  44,28,  62,39,  79,64,  117,96
New anchors saved to model. Update model *.yaml to use these anchors in the future.

函数在train.py的line206 check_anchors函数。check_anchors函数在utils/general.py。

Yolov5原本在模型配置文件（如yolov5l.py）中有默认的anchors，这些anchors是基于COCO数据集在640×640图像大小下锚定框的尺寸。Yolov5会自动按照新的数据集的labels自动学习anchors的尺寸。采用 k 均值和遗传学习算法对自定义数据集进行分析，获得适合自定义数据集中对象边界框预测的预设锚定框。

一开始会先计算Best Possible Recall (BPR)

再在kmean_anchors函数中进行k 均值和遗传学习算法更新anchors。

版权声明：本文为CSDN博主「zicai_jiayou」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/zicai_jiayou/article/details/109626712