东方耀AI技术分享
标题:
总结:SSD标签输入的编码过程:
[打印本页]
作者:
东方耀
时间:
2019-12-2 17:38
标题:
总结:SSD标签输入的编码过程:
总结:SSD标签输入的编码过程:
1、生成一个模板矩阵 shape=(batch_size, n_boxes_total, 18) 18=6+4+4+4
类别为6个0,边界框=anchor box(已经归一化),anchor box(已经归一化了), variances全为1.0
2、把模板矩阵的分类的one-hot 全部设为背景(默认)
3、一张图片一张图片的处理 人工标注的gt box标签(先归一化,转变坐标形式corners2centroids,类别进行one-hot编码)
4、当前图片中的所有人工标注的box与该张图片中所有的anchor box计算iou值
5、难点:根据iou值匹配寻找所有正样本(1对1匹配+多匹配)多匹配的阈值是设定的正样本的iou阈值
6、关键:在模板矩阵中给所有正样本的类别和边界框赋值(用人工标注框的类别和box定位坐标)
7、根据负样本的最大iou阈值限制后,找出所有的中立样本 丢弃
8、剩下就全部是负样本了:总样本数-正样本数-中立样本数
9、关键:所有图片都处理完后:将模板矩阵中的边界框坐标(已经归一化了)由绝对坐标变成 相对于anchor box的坐标,
这一步导致我们模型预测的就是相对于anchor box的坐标
10、只有正样本转换的才有意义,负样本和中立样本转换后boxes都是0(没有物体目标啊) 这种转换策略的聪明之处!
先生成一个模板y_encoded临时的(类别全0,边界框=anchor box(已经归一化了), anchor box(已经归一化了), variances全为1.): (2, 17538, 18) [[0. 0. 0. 0. 0. 0.
0.00833333 0.01351351 0.03535534 0.11313708 0.00833333 0.01351351
0.03535534 0.11313708 1. 1. 1. 1. ]
[0. 0. 0. 0. 0. 0.
0.00833333 0.01351351 0.02886751 0.13856406 0.00833333 0.01351351
0.02886751 0.13856406 1. 1. 1. 1. ]
[0. 0. 0. 0. 0. 0.
0.00833333 0.01351351 0.05 0.08 0.00833333 0.01351351
0.05 0.08 1. 1. 1. 1. ]]
把模板的y_encoded分类的one-hot 全部设为背景
正在处理的图像编号:0
该图片的人工标注框labels: (5, 5) [[ 1. 87. 153. 123. 173.]
[ 1. 146. 153. 164. 167.]
[ 1. 190. 151. 219. 174.]
[ 1. 263. 143. 299. 176.]
[ 1. 295. 140. 350. 185.]]
该图片的人工标注框(归一化后,坐标从corners变为中心点(cx cy w h)形式): [[1. 0.21875 0.54333333 0.075 0.06666667]
[1. 0.32291667 0.53333333 0.0375 0.04666667]
[1. 0.42604167 0.54166667 0.06041667 0.07666667]
[1. 0.58541667 0.53166667 0.075 0.11 ]
[1. 0.671875 0.54166667 0.11458333 0.15 ]]
该图片的人工标注框(归一化后,坐标从corners变为中心点(cx cy w h)形式,类别one-hot编码后)labels_one_hot: (5, 10) [[0. 1. 0. 0. 0. 0.
0.21875 0.54333333 0.075 0.06666667]
[0. 1. 0. 0. 0. 0.
0.32291667 0.53333333 0.0375 0.04666667]
[0. 1. 0. 0. 0. 0.
0.42604167 0.54166667 0.06041667 0.07666667]
[0. 1. 0. 0. 0. 0.
0.58541667 0.53166667 0.075 0.11 ]
[0. 1. 0. 0. 0. 0.
0.671875 0.54166667 0.11458333 0.15 ]]
该图片中的所有人工标注的box与该张图片中所有的anchor box计算iou值(都归一化了 坐标格式都是中心点的):similarities (5, 17538) [[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]]
严格匹配,最值且唯一不重复的匹配(最优的一夫一妻制):anchor box的索引: [ 7283 6956 7352 7053 15062]
已经匹配上的anchor box就不能再去匹配了,类似找女朋友一样
除开与gt box最优匹配的anchor box之外,剩下的anchor box都可以主动去找之最大匹配的gt box(肯定会重复)经过正样本的iou阈值的筛选后 返回gt_indices: 22 [0 0 2 2 2 3 0 0 0 0 0 0 2 2 3 3 3 4 4 4 4 4]
返回的anchor_indices: 22 [ 6921 6922 6992 6993 6994 7047 7274 7276 7277 7280 7281 7282
7353 7354 7407 7413 15044 15056 15058 15063 15064 15243]
以上过程:已经把正样本全部找出来了,正样本的总个数: 27
计算iou的similarities shape不变: (5, 17538)
max_background_similarities(除了正样本后还有多少iou值不为0)(看前100个): 2337 [0.00776545 0.02395913 0.02794035 0.02794035 0.00437515 0.00478032
0.02794035 0.01029781 0.01183467 0.02591161 0.01629073 0.01493905
0.03022686 0.02235518 0.01493905 0.03022686 0.02539204 0.01493905
0.03022686 0.02539204 0.01493905 0.03022686 0.02539204 0.01493905
0.03022686 0.02007255 0.00917771 0.01926914 0.01403507 0.00216023
0.00196928 0.00806864 0.00217201 0.00477654 0.0146467 0.01705473
0.01705473 0.01705473 0.00722603 0.00176129 0.01720269 0.0269542
0.01329788 0.01435088 0.03854473 0.04142676 0.04142676 0.01731328
0.0015806 0.02018475 0.02470847 0.02644253 0.01440761 0.04854334
0.06301532 0.08486428 0.02756742 0.0785236 0.0806146 0.09995386
0.03936556 0.11026872 0.00041663 0.0806146 0.09995386 0.03936556
0.11849917 0.00446018 0.00228858 0.07304377 0.09995386 0.0309109
0.08628847 0.01049916 0.00416755 0.04990594 0.06039665 0.03330008
0.06968401 0.01661119 0.00605357 0.06362481 0.07085155 0.04766545
0.09753235 0.02279761 0.00794669 0.06362481 0.07085155 0.04766545
0.11189953 0.0258961 0.00979569 0.06362481 0.07085155 0.04766545
0.11189953 0.0258961 0.00979569 0.06362481]
经过负样本的最大阈值限制后,找出所有的中立样本neutral_boxes个数: 187 [ 6561 6594 6595 6633 6686 6687 6688 6689 6692 6693 6694 6695
6699 6717 6723 6729 6909 6911 6912 6914 6915 6916 6917 6918
6920 6923 6924 6926 6927 6928 6929 6950 6952 6953 6954 6955
6958 6959 6964 6965 6984 6985 6986 6987 6988 6989 6990 6991
6995 6996 6997 6998 6999 7000 7001 7038 7040 7041 7044 7045
7046 7048 7049 7050 7051 7052 7054 7055 7056 7057 7058 7059
7071 7077 7083 7089 7095 7268 7269 7270 7271 7272 7275 7278
7284 7286 7287 7288 7289 7314 7315 7316 7344 7345 7346 7347
7348 7349 7350 7351 7355 7356 7357 7358 7359 7360 7361 7401
7404 7405 7406 7408 7409 7410 7411 7412 7414 7415 7416 7417
7418 7419 7431 7437 7443 7449 7455 7635 7641 7713 7767 7773
7797 7803 7809 14862 14863 14864 14874 14875 14877 14880 14881 14882
14883 14974 14978 14980 14986 15038 15040 15041 15042 15043 15046 15047
15050 15052 15053 15054 15055 15057 15059 15060 15061 15065 15068 15069
15070 15071 15158 15222 15223 15224 15234 15235 15236 15237 15238 15239
15240 15241 15242 15244 15245 15249 15251]
所有负样本的个数:(y_encoded中) 17324
中立样本(y_encoded中分类的one-hot全为0)个数: 187
所有负样本的个数17324:= (总样本数17538-正样本数27-中立样本数187)
正在处理的图像编号:1
该图片的人工标注框labels: (19, 5) [[ 1. 36. 161. 60. 170.]
[ 1. 50. 159. 70. 168.]
[ 1. 105. 156. 123. 165.]
[ 1. 106. 157. 125. 169.]
[ 1. 124. 150. 150. 170.]
[ 1. 158. 154. 165. 159.]
[ 1. 170. 150. 190. 166.]
[ 1. 185. 150. 198. 160.]
[ 1. 196. 150. 222. 170.]
[ 1. 259. 146. 294. 159.]
[ 2. 64. 155. 82. 165.]
[ 5. 80. 139. 85. 149.]
[ 5. 80. 140. 86. 152.]
[ 5. 151. 118. 157. 129.]
[ 5. 153. 119. 159. 130.]
[ 5. 180. 117. 187. 127.]
[ 5. 182. 117. 186. 126.]
[ 5. 230. 134. 236. 143.]
[ 5. 230. 135. 235. 145.]]
该图片的人工标注框(归一化后,坐标从corners变为中心点(cx cy w h)形式): [[1. 0.1 0.55166667 0.05 0.03 ]
[1. 0.125 0.545 0.04166667 0.03 ]
[1. 0.2375 0.535 0.0375 0.03 ]
[1. 0.240625 0.54333333 0.03958333 0.04 ]
[1. 0.28541667 0.53333333 0.05416667 0.06666667]
[1. 0.33645833 0.52166667 0.01458333 0.01666667]
[1. 0.375 0.52666667 0.04166667 0.05333333]
[1. 0.39895833 0.51666667 0.02708333 0.03333333]
[1. 0.43541667 0.53333333 0.05416667 0.06666667]
[1. 0.57604167 0.50833333 0.07291667 0.04333333]
[2. 0.15208333 0.53333333 0.0375 0.03333333]
[5. 0.171875 0.48 0.01041667 0.03333333]
[5. 0.17291667 0.48666667 0.0125 0.04 ]
[5. 0.32083333 0.41166667 0.0125 0.03666667]
[5. 0.325 0.415 0.0125 0.03666667]
[5. 0.38229167 0.40666667 0.01458333 0.03333333]
[5. 0.38333333 0.405 0.00833333 0.03 ]
[5. 0.48541667 0.46166667 0.0125 0.03 ]
[5. 0.484375 0.46666667 0.01041667 0.03333333]]
该图片的人工标注框(归一化后,坐标从corners变为中心点(cx cy w h)形式,类别one-hot编码后)labels_one_hot: (19, 10) [[0. 1. 0. 0. 0. 0.
0.1 0.55166667 0.05 0.03 ]
[0. 1. 0. 0. 0. 0.
0.125 0.545 0.04166667 0.03 ]
[0. 1. 0. 0. 0. 0.
0.2375 0.535 0.0375 0.03 ]
[0. 1. 0. 0. 0. 0.
0.240625 0.54333333 0.03958333 0.04 ]
[0. 1. 0. 0. 0. 0.
0.28541667 0.53333333 0.05416667 0.06666667]
[0. 1. 0. 0. 0. 0.
0.33645833 0.52166667 0.01458333 0.01666667]
[0. 1. 0. 0. 0. 0.
0.375 0.52666667 0.04166667 0.05333333]
[0. 1. 0. 0. 0. 0.
0.39895833 0.51666667 0.02708333 0.03333333]
[0. 1. 0. 0. 0. 0.
0.43541667 0.53333333 0.05416667 0.06666667]
[0. 1. 0. 0. 0. 0.
0.57604167 0.50833333 0.07291667 0.04333333]
[0. 0. 1. 0. 0. 0.
0.15208333 0.53333333 0.0375 0.03333333]
[0. 0. 0. 0. 0. 1.
0.171875 0.48 0.01041667 0.03333333]
[0. 0. 0. 0. 0. 1.
0.17291667 0.48666667 0.0125 0.04 ]
[0. 0. 0. 0. 0. 1.
0.32083333 0.41166667 0.0125 0.03666667]
[0. 0. 0. 0. 0. 1.
0.325 0.415 0.0125 0.03666667]
[0. 0. 0. 0. 0. 1.
0.38229167 0.40666667 0.01458333 0.03333333]
[0. 0. 0. 0. 0. 1.
0.38333333 0.405 0.00833333 0.03 ]
[0. 0. 0. 0. 0. 1.
0.48541667 0.46166667 0.0125 0.03 ]
[0. 0. 0. 0. 0. 1.
0.484375 0.46666667 0.01041667 0.03333333]]
该图片中的所有人工标注的box与该张图片中所有的anchor box计算iou值(都归一化了 坐标格式都是中心点的):similarities (19, 17538) [[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0.]]
严格匹配,最值且唯一不重复的匹配(最优的一夫一妻制):anchor box的索引: [7241 7246 6935 6926 6944 6953 6976 6982 6998 6688 6905 5820 6180 4795
5507 5174 4813 5575 6287]
已经匹配上的anchor box就不能再去匹配了,类似找女朋友一样
除开与gt box最优匹配的anchor box之外,剩下的anchor box都可以主动去找之最大匹配的gt box(肯定会重复)经过正样本的iou阈值的筛选后 返回gt_indices: 9 [9 4 4 4 6 6 8 8 8]
返回的anchor_indices: 9 [6689 6938 6940 6946 6970 6974 6992 6994 7000]
以上过程:已经把正样本全部找出来了,正样本的总个数: 28
计算iou的similarities shape不变: (19, 17538)
max_background_similarities(除了正样本后还有多少iou值不为0)(看前100个): 2335 [0.0015805 0.00275804 0.02596375 0.0015883 0.00301762 0.04020804
0.0015883 0.00074531 0.0095082 0.0015883 0.00099402 0.00763012
0.01453231 0.05829291 0.00763012 0.01453231 0.04946288 0.00763012
0.00763012 0.00541776 0.0431047 0.07761704 0.07216262 0.03341164
0.04332643 0.08552322 0.03341164 0.04332643 0.04203139 0.03416654
0.02768885 0.04332643 0.00142141 0.00713446 0.02839687 0.05710271
0.01034301 0.11381261 0.05077593 0.05710271 0.01034301 0.11381261
0.10104008 0.05077593 0.05710271 0.01034301 0.02114752 0.05710271
0.01034301 0.01838876 0.01743265 0.05699456 0.08095374 0.06463182
0.10373396 0.07216262 0.11458333 0.05729167 0.08138535 0.06463182
0.11458333 0.11458333 0.11458333 0.05729167 0.08138535 0.06463182
0.06350922 0.03416654 0.1095852 0.05729167 0.08138535 0.06463182
0.00756895 0.0660066 0.06076389 0.10876532 0.08840401 0.12152778
0.12013157 0.06076389 0.10876532 0.08840401 0.12152778 0.10104008
0.12152778 0.06076389 0.10876532 0.08840401 0.0487013 0.06076389
0.10876532 0.08840401 0.04593122 0.00268445 0.00515145 0.03524502
0.00268445 0.00515145 0.04324164 0.00268445]
经过负样本的最大阈值限制后,找出所有的中立样本neutral_boxes个数: 115 [6327 6576 6577 6579 6582 6583 6584 6585 6612 6613 6614 6630 6631 6633
6636 6637 6638 6639 6677 6680 6682 6683 6686 6687 6692 6694 6695 6701
6884 6892 6893 6896 6898 6899 6924 6934 6936 6937 6939 6941 6942 6943
6945 6947 6950 6952 6968 6971 6972 6973 6977 6980 6983 6988 6989 6990
6991 6993 6995 6996 6997 6999 7001 7004 7006 7007 7040 7042 7046 7047
7048 7049 7052 7054 7055 7229 7234 7235 7240 7244 7247 7256 7282 7284
7286 7288 7294 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306
7307 7332 7333 7334 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359
7360 7361 7644]
所有负样本的个数:(y_encoded中) 17395
中立样本(y_encoded中分类的one-hot全为0)个数: 115
所有负样本的个数17395:= (总样本数17538-正样本数28-中立样本数115)
所有图片都处理完后:将边界框坐标(已经归一化了)格式是中心点形式 由绝对坐标变成 相对于anchor box的坐标:
只看正样本转换的才有意义,负样本和中立样本转换后boxes都是0(没有物体目标啊),这才是:这样转换的核心好处
看看转换之前的(某个正样本): [6921] [[0. 1. 0. 0. 0. 0.
0.21875 0.54333333 0.075 0.06666667 0.225 0.52702703
0.07071068 0.11313708 1. 1. 1. 1. ]]
看看转换之前的(某个负样本): [0] [[1. 0. 0. 0. 0. 0.
0.00833333 0.01351351 0.03535534 0.11313708 0.00833333 0.01351351
0.03535534 0.11313708 1. 1. 1. 1. ]]
看看转换之前的(某个中立样本): [6561] [[0. 0. 0. 0. 0. 0.
0.225 0.5 0.07071068 0.11313708 0.225 0.5
0.07071068 0.11313708 1. 1. 1. 1. ]]
转换之后的(某个正样本): [6921] [[ 0. 1. 0. 0. 0. 0.
-0.08838835 0.14412875 0.05889152 -0.52889515 0.225 0.52702703
0.07071068 0.11313708 1. 1. 1. 1. ]]
转换之后的(某个负样本): [0] [[1. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0.00833333 0.01351351
0.03535534 0.11313708 1. 1. 1. 1. ]]
转换之后的(某个中立样本): [6561] [[0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0.225 0.5
0.07071068 0.11313708 1. 1. 1. 1. ]]
复制代码
作者:
xsoft
时间:
2020-2-3 15:36
谢谢老师提供的资料。
欢迎光临 东方耀AI技术分享 (http://www.ai111.vip/)
Powered by Discuz! X3.4