人脸检测 Retinaface - GT框（ground truth box）与 先验框（prior box）之间是如何匹配的
人脸检测 Retinaface - GT框（ground truth box）与 先验框（prior box）之间是如何匹配的
flyfish
原函数是def match(threshold, truths, priors, variances, labels, landms, loc_t, conf_t, landm_t, idx):
内容较多所以将函数简化重新解释
开始部分是计算IoU
计算IoU 是计算先验框和GT框
先验框的坐标是center-size，这里要转换成统一的 xmin, ymin, xmax, ymax形式
truths是GT框的坐标，priors是先验框的坐标，计算前要用point_form转换坐标，所以看到的代码是
overlaps = jaccard(truths,point_form(priors) )
后面的操作模拟写一个程序
先把数量边小,手工标注一张图片，里面有2个目标，程序自动生成了5个先验框，然后分析
import torch
print( torch.__version__)#1.4.0
truths=torch.randn(2,4)#[目标，每个边框坐标4个值]
labels=torch.Tensor([[99], [100]])#两个类别，类别数据比较大，容易一眼就看出来
overlaps = torch.rand(2, 5)#[目标的个数，先验框的个数] 生成0到1之间的随机数作为IoU的计算结果
print("overlaps:",overlaps)
# overlaps: \
# tensor([[0.8388, 0.9715, 0.4808, 0.4071, 0.2703],
# [0.1282, 0.7579, 0.6970, 0.3630, 0.5478]])
#为每个GT框匹配一个IoU最大的先验框，获取 与目标GT框（GTBox） IoU最大的 先验框（PriorBox） 的值和索引
#GTBox 1：PriorBox n
best_prior_overlap, best_prior_idx = overlaps.max(1, keepdim=True)
print(overlaps.max(1, keepdim=True))
# torch.return_types.max(
# values=tensor([[0.9715],
# [0.7579]]),
# indices=tensor([[1],
# [1]]))
#与上面相反，获取 与先验框（PriorBox） IoU最大的 GT框（GTBox） 的值和索引
#PriorBox 1：GTBox n
best_truth_overlap, best_truth_idx = overlaps.max(0, keepdim=True)
print(overlaps.max(0, keepdim=True))
# torch.return_types.max(
# values=tensor([[0.8388, 0.9715, 0.6970, 0.4071, 0.5478]]),
# indices=tensor([[0, 0, 1, 0, 1]]))
best_prior_idx.squeeze_(1) # best_prior_idx的shape是[num_objects,1],去掉1维，shape变为[num_objects]
best_prior_overlap.squeeze_(1) # best_prior_idx的shape是[num_objects,1],去掉1维，shape变为[num_objects]
best_truth_idx.squeeze_(0) # best_truth_idx的shape是[1,num_priors],去掉0维，shape变为[num_priors]
best_truth_overlap.squeeze_(0) # best_truth_overlap的shape是[1,num_priors],去掉0维，shape变为[num_priors]
#GTBox与PriorBox的IoU最大的，确保最好的PriorBox保留下来，设置为2,只要大于阈值就行，设置88也没问题.
best_truth_overlap.index_fill_(0, best_prior_idx, 88)
print("best_truth_overlap:",best_truth_overlap)
# best_truth_overlap: \
# tensor([ 0.8388, 88.0000, 0.6970, 0.4071, 0.5478])
# 保证每一个GTBox匹配它的都是具有最大IoU的PriorBox
for j in range(best_prior_idx.size(0)):
best_truth_idx[best_prior_idx[j]] = j
# 循环过程
# best_prior_idx.size(0): 2
# j 0
# best_prior_idx[j]: tensor(1)
# j 1
# best_prior_idx[j]: tensor(1)
# best_truth_idx: tensor([0, 1, 1, 0, 1])
print("best_truth_idx:",best_truth_idx)
#best_truth_idx: tensor([0, 1, 1, 0, 1])
matches = truths[best_truth_idx]#每一个PriorBox对应的bbox取出来
print("matches:",matches)
# matches: tensor([[ 0.8280, 0.5840, 0.5258, 1.1107],
# [-1.4758, 0.3446, 0.3785, 1.8465],
# [-1.4758, 0.3446, 0.3785, 1.8465],
# [ 0.8280, 0.5840, 0.5258, 1.1107],
# [-1.4758, 0.3446, 0.3785, 1.8465]])
print("best_truth_overlap：",best_truth_overlap)#每一个anchor对应的label取出来
#best_truth_overlap： tensor([ 0.8388, 88.0000, 0.6970, 0.4071, 0.5478])
conf = labels[best_truth_idx] #PriorBox对应的gt box的类别
print(conf.shape)#torch.Size([5, 1])
print("conf:",conf)
# conf: tensor([[ 99.],
# [100.],
# [100.],
# [ 99.],
# [100.]])
threshold=0.1
conf[best_truth_overlap < threshold] = 0 #过滤掉iou太低的,标记为background
print("conf:",conf)
# conf: tensor([[ 99.],
# [100.],
# [100.],
# [ 99.],
# [100.]])
相当于遍历conf中每一个元素都与threshold比较，成立则该元素为0
假设threshold=0.5
假设best_truth_overlap的值是
tensor([ 0.2567, 0.8579, 0.3156, 88.0000, 88.0000])
结果是
# conf: tensor([[100.],
# [ 99.],
# [100.],
# [100.],
# [ 99.]])
# 则过滤掉其中的0.2567和 0.3156
# conf: tensor([[ 0.],
# [ 99.],
# [ 0.],
# [100.],
# [ 99.]])