人臉檢測 Retinaface - GT框（ground truth box）與 先驗框（prior box）之間是如何匹配的
flyfish
原函數是def match(threshold, truths, priors, variances, labels, landms, loc_t, conf_t, landm_t, idx):
內容較多所以將函數簡化重新解釋
開始部分是計算IoU
計算IoU 是計算先驗框和GT框
先驗框的座標是center-size，這裏要轉換成統一的 xmin, ymin, xmax, ymax形式
truths是GT框的座標，priors是先驗框的座標，計算前要用point_form轉換座標，所以看到的代碼是
overlaps = jaccard(truths,point_form(priors) )
後面的操作模擬寫一個程序
先把數量邊小,手工標註一張圖片，裏面有2個目標，程序自動生成了5個先驗框，然後分析
import torch
print( torch.__version__)#1.4.0
truths=torch.randn(2,4)#[目標，每個邊框座標4個值]
labels=torch.Tensor([[99], [100]])#兩個類別，類別數據比較大，容易一眼就看出來
overlaps = torch.rand(2, 5)#[目標的個數，先驗框的個數] 生成0到1之間的隨機數作爲IoU的計算結果
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]])
#爲每個GT框匹配一個IoU最大的先驗框，獲取 與目標GT框（GTBox） IoU最大的 先驗框（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]]))
#與上面相反，獲取 與先驗框（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維，shape變爲[num_objects]
best_prior_overlap.squeeze_(1) # best_prior_idx的shape是[num_objects,1],去掉1維，shape變爲[num_objects]
best_truth_idx.squeeze_(0) # best_truth_idx的shape是[1,num_priors],去掉0維，shape變爲[num_priors]
best_truth_overlap.squeeze_(0) # best_truth_overlap的shape是[1,num_priors],去掉0維，shape變爲[num_priors]
#GTBox與PriorBox的IoU最大的，確保最好的PriorBox保留下來，設置爲2,只要大於閾值就行，設置88也沒問題.
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])
# 保證每一個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
# 循環過程
# 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]#每一個PriorBox對應的bbox取出來
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)#每一個anchor對應的label取出來
#best_truth_overlap： tensor([ 0.8388, 88.0000, 0.6970, 0.4071, 0.5478])
conf = labels[best_truth_idx] #PriorBox對應的gt box的類別
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 #過濾掉iou太低的,標記爲background
print("conf:",conf)
# conf: tensor([[ 99.],
# [100.],
# [100.],
# [ 99.],
# [100.]])
相當於遍歷conf中每一個元素都與threshold比較，成立則該元素爲0
假設threshold=0.5
假設best_truth_overlap的值是
tensor([ 0.2567, 0.8579, 0.3156, 88.0000, 88.0000])
結果是
# conf: tensor([[100.],
# [ 99.],
# [100.],
# [100.],
# [ 99.]])
# 則過濾掉其中的0.2567和 0.3156
# conf: tensor([[ 0.],
# [ 99.],
# [ 0.],
# [100.],
# [ 99.]])