anchor boxes用来预测bounding box,faster rcnn中用128*128,256*256,512*512,分三个尺度变换1:1,1:2,2:1,共计9个anchor来预测框,每个anchor预测2000个框左右,使得检出率提高很多。YOLOv2开始增加了anchor机制,在v3中增加到9个anchor。例如yolov3-voc.cfg中这组anchor,anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326,由作者通过聚类VOC数据集得到的,20类目标中大到bicycle、bus,小到bird、cat,目标大小差距很大,如果用自己的数据集训练检测目标,其中部分anchor并不合理,本文记录下在自己的数据集上聚类计算anchor,提高bounding box的检出率。
原工程:https://github.com/lars76/kmeans-anchor-boxes
Joseph Redmon论文数据avg iou在67.2,该作者验证在k=9时,多次迭代在VOC 2007数据集上得到avg iou在67.13,相差无几。
修改的example.py
import glob
import xml.etree.ElementTree as ET
import numpy as np
from kmeans import kmeans, avg_iou
ANNOTATIONS_PATH = "Annotations"
CLUSTERS = 9
def load_dataset(path):
dataset = []
for xml_file in glob.glob("{}/*xml".format(path)):
tree = ET.parse(xml_file)
height = int(tree.findtext("./size/height"))
width = int(tree.findtext("./size/width"))
for obj in tree.iter("object"):
xmin = int(obj.findtext("bndbox/xmin")) / width
ymin = int(obj.findtext("bndbox/ymin")) / height
xmax = int(obj.findtext("bndbox/xmax")) / width
ymax = int(obj.findtext("bndbox/ymax")) / height
xmin = np.float64(xmin)
ymin = np.float64(ymin)
xmax = np.float64(xmax)
ymax = np.float64(ymax)
if xmax == xmin or ymax == ymin:
print(xml_file)
dataset.append([xmax - xmin, ymax - ymin])
return np.array(dataset)
if __name__ == '__main__':
#print(__file__)
data = load_dataset(ANNOTATIONS_PATH)
out = kmeans(data, k=CLUSTERS)
#clusters = [[10,13],[16,30],[33,23],[30,61],[62,45],[59,119],[116,90],[156,198],[373,326]]
#out= np.array(clusters)/416.0
print(out)
print("Accuracy: {:.2f}%".format(avg_iou(data, out) * 100))
print("Boxes:\n {}-{}".format(out[:, 0]*416, out[:, 1]*416))
ratios = np.around(out[:, 0] / out[:, 1], decimals=2).tolist()
print("Ratios:\n {}".format(sorted(ratios)))
Kmeans因为初始点敏感,所以每次运行得到的anchor值不一样,但是对应的avg iou稳定。用于训练的话就需要统计多组anchor,针对固定的测试集比较了。
可以计算下VOC的这组anchor在自己数据集上的avg iou,对比直接在数据集上聚类得到的anchor以及avg iou。
参考:
https://blog.csdn.net/hrsstudy/article/details/71173305?utm_source=itdadao&utm_medium=referral
https://blog.csdn.net/sinat_24143931/article/details/78773936