-
Notifications
You must be signed in to change notification settings - Fork 2
/
metric.py
308 lines (245 loc) · 10.7 KB
/
metric.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
import numpy as np
np.set_printoptions(suppress=True)
def compute_iou(bboxes1, bboxes2, bbox_format='cxcywh'):
if bbox_format == 'cxcywh':
# if bboxes are composed of center point xy and wh/2.
cx1, cy1, w1, h1 = np.split(bboxes1, 4, axis=1)
cx2, cy2, w2, h2 = np.split(bboxes2, 4, axis=1)
x11 = cx1 - w1 / 2
y11 = cy1 - h1 / 2
x12 = cx1 + w1 / 2
y12 = cy1 + h1 / 2
x21 = cx2 - w2 /2
y21 = cy2 - h2 /2
x22 = cx2 + w2 /2
y22 = cy2 + h2 /2
elif bbox_format == 'tlxtlywh':
# if bboxes are composed of top left point xy and wh.
tlx1, tly1, w1, h1 = np.split(bboxes1, 4, axis=1)
tlx2, tly2, w2, h2 = np.split(bboxes2, 4, axis=1)
x11 = tlx1
y11 = tly1
x12 = tlx1 + w1
y12 = tly1 + h1
x21 = tlx2
y21 = tly2
x22 = tlx2 + w2
y22 = tly2 + h2
else:
# if bboxes are composed of top left point xy and bottom right xy.
x11, y11, x12, y12 = np.split(bboxes1, 4, axis=1)
x21, y21, x22, y22 = np.split(bboxes2, 4, axis=1)
# determine the (x, y)-coordinates of the intersection rectangle
inter_x1 = np.maximum(x11, x21.T)
inter_y1 = np.maximum(y11, y21.T)
inter_x2 = np.minimum(x12, x22.T)
inter_y2 = np.minimum(y12, y22.T)
# # compute the area of intersection rectangle
inter_area = np.maximum(inter_x2 - inter_x1 + 1 , 1e-6) * np.maximum(inter_y2 - inter_y1 + 1 , 1e-6)
# # compute the area of both the prediction and ground-truth rectangles
bboxes1_area = (x12 - x11 + 1 ) * (y12 - y11 + 1)
bboxes2_area = (x22 - x21 + 1 ) * (y22 - y21 + 1)
# compute the area of intersection rectangle
# inter_area = np.maximum(inter_x2 - inter_x1, 0) * np.maximum(inter_y2 - inter_y1, 0)
# # compute the area of both the prediction and ground-truth rectangles
# bboxes1_area = (x12 - x11) * (y12 - y11)
# bboxes2_area = (x22 - x21) * (y22 - y21)
iou = inter_area / ((bboxes1_area + bboxes2_area.T - inter_area) + 1e-5)
assert np.min(iou) >= 0
return iou
def measure_tpfp(pred_bboxes, gt_bboxes, iou_threshold=0.5, bbox_format='cxcywh'):
'''
https://github.com/rafaelpadilla/Object-Detection-Metrics
위 프로젝트의 https://github.com/rafaelpadilla/Object-Detection-Metrics/blob/master/pascalvoc.py
코드 및 voc 평가방법 참고
pred_bboxes: shape:[num_pred_bboxes, 6], class, x, y, w, h, confidence score
gt_bboxes: shape:[num_gt_bboxes, 5], class, x, y, w, h
iou_threshold: 예측된 바운딩 박스를 tp/fp로 분류하기위한 iou thereshold
return class_tp_fp_score
class_tp_fp_score: shape:[num_pred_bboxes, 4], class, tp, fp, confidence score
예측된 바운딩 박스별로 class tp/fp, confidence score 값을 기록한 table을 return 함
'''
pred_bboxes = pred_bboxes[np.argsort(pred_bboxes[:, 5])[::-1]]
# print(pred_bboxes)
# exit()
iou_per_box = np.zeros((len(pred_bboxes), 1))
tp_per_box = np.zeros((len(pred_bboxes), 1))
for c in np.unique(gt_bboxes[:, 0]):
gt_mask = gt_bboxes[:, 0] == c
pred_mask = pred_bboxes[:, 0] == c
filtered_gt_bboxes = gt_bboxes[gt_mask] # (M, )
filtered_pred_bboxes = pred_bboxes[pred_mask] # (N, )
if len(filtered_gt_bboxes) == 0 or len(filtered_pred_bboxes) == 0:
continue
filtered_iou_per_box = iou_per_box[pred_mask]
filtered_tp_per_box = tp_per_box[pred_mask]
iou_matrix = compute_iou(filtered_pred_bboxes[:, 1:5], filtered_gt_bboxes[:, 1:5], bbox_format) # (N, M)
#for iou in iou_matrix:
#print("np.sum(iou_matrix): ", np.sum(iou_matrix))
# TP 여부를 판단할때, 어떤 한 Predicted Box가 두 Ground Truth Box에 대한 검출 결과로 중복적으로 매칭될 수 있다.
# 이 경우 다른 Predicted Box가 두 Ground Truth Box중 하나의 Box에 대한 검출 결과로 매칭되고 TP가 될 가능성이 생긴다.
# 다만 VOC 는 이런 경우를 신경쓰지않음...
for i in range(len(filtered_gt_bboxes)):
iou = iou_matrix[:, i] # filtered_pred_bboxes와 i번째 filtered_gt_bbox간 IoU
matched = False
for j in range(len(iou)):
if filtered_tp_per_box[j] == 1.: # already matched
continue
if iou[j] >= iou_threshold:
if not matched:
matched = True
filtered_iou_per_box[j] = iou[j]
filtered_tp_per_box[j] = 1.
iou_per_box[pred_mask] = filtered_iou_per_box
tp_per_box[pred_mask] = filtered_tp_per_box
fp_per_box = 1 - tp_per_box
# print(pred_bboxes.shape)
# print(np.sum(tp_per_box) + np.sum(fp_per_box))
assert np.sum(tp_per_box) <= len(gt_bboxes), "Your code is wrong. The number of TP cases cannot exceed the number of ground truth boxes."
assert np.sum(tp_per_box) + np.sum(fp_per_box) <= len(pred_bboxes), "Your code is wrong"
# print(gt_bboxes)
# print(fp_per_box)
# print("--------------------")
# exit()
class_tp_fp_score = np.concatenate((pred_bboxes[:, 0, None],
tp_per_box,
fp_per_box,
pred_bboxes[:, 5, None]),
axis=1)#(N, 3)
sorted_inds_by_score = np.argsort(pred_bboxes[:, 5])[::-1]#내림차순
class_tp_fp_score = class_tp_fp_score[sorted_inds_by_score]
return class_tp_fp_score
def compute_ap(tp_fp_score, gt_bboxes):
tp_fp_score = tp_fp_score[np.argsort(tp_fp_score[:, -1])[::-1]]
num_all_gt_bboxes = len(gt_bboxes)
accumulated_tp = np.cumsum(tp_fp_score[:, 0]).reshape(-1, 1)
accumulated_fp = np.cumsum(tp_fp_score[:, 1]).reshape(-1, 1)
precision = accumulated_tp/(accumulated_tp + accumulated_fp)
recall = accumulated_tp/num_all_gt_bboxes
# print(precision)
# import matplotlib.pyplot as plt
# plt.plot(recall, precision)
# plt.show()
#return CalculateAveragePrecision(recall, precision)[:3]
interpolated_precision = np.zeros_like(precision)
for i in range(len(precision)):
interpolated_precision[i] = np.max(precision[i:])
recall = list(recall)
interpolated_precision = list(interpolated_precision)
recall.insert(0, 0)
interpolated_precision.insert(0, interpolated_precision[0])
ap = 0.
for i in range(1, len(recall)):
ap = ap + interpolated_precision[i] * (recall[i] - recall[i - 1])
return ap, precision, recall
def compute_map(class_tp_fp_score, gt_bboxes, num_classes):
class_tp_fp_score = np.concatenate(class_tp_fp_score, axis=0)
class_tp_fp_score = class_tp_fp_score[np.argsort(class_tp_fp_score[:, 3])[::-1]]
gt_bboxes = np.concatenate(gt_bboxes, axis=0)
ap_per_class = np.zeros(num_classes)
for c in range(num_classes):
if np.count_nonzero(class_tp_fp_score[:, 0] == c) == 0:
continue
if np.count_nonzero(gt_bboxes[:, 0] == c) == 0:
continue
ap, _, _ = compute_ap(class_tp_fp_score[class_tp_fp_score[:, 0] == c, 1:], gt_bboxes[gt_bboxes[:, 0] == c, 1:])
ap_per_class[c] = ap
print(ap_per_class)
mean_ap = np.mean(ap_per_class)
return mean_ap
if __name__ == '__main__':
num_classes = 1
pred_bboxes_batch = []
gt_bboxes_batch =[]
#xywh
pred_bboxes1 = np.array([
[ 0 , 5.0 , 67.0 , 31.0 , 48.0 , 0.88 ],
[ 0 , 119.0 , 111.0 , 40.0 , 67.0 , 0.7 ],
[ 0 , 124.0 , 9.0 , 49.0 , 67.0 , 0.8 ]
])
pred_bboxes2 = np.array([
[ 0 , 64.0 , 111.0 , 64.0 , 58.0 , 0.71 ],
[ 0 , 26.0 , 140.0 , 60.0 , 47.0 , 0.54 ],
[ 0 , 19.0 , 18.0 , 43.0 , 35.0 , 0.74 ]
])
pred_bboxes3 = np.array([
[ 0 , 109.0 , 15.0 , 77.0 , 39.0 , 0.18 ],
[ 0 , 86.0 , 63.0 , 46.0 , 45.0 , 0.67 ],
[ 0 , 160.0 , 62.0 , 36.0 , 53.0 , 0.38 ],
[ 0 , 105.0 , 131.0 , 47.0 , 47.0 , 0.91 ],
[ 0 , 18.0 , 148.0 , 40.0 , 44.0 , 0.44 ]
])
pred_bboxes4 = np.array([
[ 0 , 83.0 , 28.0 , 28.0 , 26.0 , 0.35 ],
[ 0 , 28.0 , 68.0 , 42.0 , 67.0 , 0.78 ],
[ 0 , 87.0 , 89.0 , 25.0 , 39.0 , 0.45 ],#망한 박스
[ 0 , 10.0 , 155.0 , 60.0 , 26.0 , 0.14 ]
])
pred_bboxes5 = np.array([
[ 0 , 50.0 , 38.0 , 28.0 , 46.0 , 0.62 ],
[ 0 , 95.0 , 11.0 , 53.0 , 28.0 , 0.44 ],
[ 0 , 29.0 , 131.0 , 72.0 , 29.0 , 0.95 ],
[ 0 , 29.0 , 163.0 , 72.0 , 29.0 , 0.23 ]
])
pred_bboxes6 = np.array([
[ 0 , 43.0 , 48.0 , 74.0 , 38.0 , 0.45 ],
[ 0 , 17.0 , 155.0 , 29.0 , 35.0 , 0.84 ],
[ 0 , 95.0 , 110.0 , 25.0 , 42.0 , 0.43 ]
])
pred_bboxes7 = np.array([
[ 0 , 16.0 , 20.0 , 101.0 , 88.0 , 0.48 ],
[ 0 , 33.0 , 116.0 , 37.0 , 49.0 , 0.95 ]
])
gt_bboxes1 = np.array([
[ 0 , 25.0 , 16.0 , 38.0 , 56.0 ],
[ 0 , 129.0 , 123.0 , 41.0 , 62.0 ]
])
gt_bboxes2 = np.array([
[ 0 , 123.0 , 11.0 , 43.0 , 55.0 ],
[ 0 , 38.0 , 132.0 , 59.0 , 45.0 ]
])
gt_bboxes3 = np.array([
[ 0 , 16.0 , 14.0 , 35.0 , 48.0 ],
[ 0 , 123.0 , 30.0 , 49.0 , 44.0 ],
[ 0 , 99.0 , 139.0 , 47.0 , 47.0 ]
])
gt_bboxes4 = np.array([
[ 0 , 53.0 , 42.0 , 40.0 , 52.0 ],
[ 0 , 154.0 , 43.0 , 31.0 , 34.0 ]
])
gt_bboxes5 = np.array([
[ 0 , 59.0 , 31.0 , 44.0 , 51.0 ],
[ 0 , 48.0 , 128.0 , 34.0 , 52.0 ]
])
gt_bboxes6 = np.array([
[ 0 , 36.0 , 89.0 , 52.0 , 76.0 ],
[ 0 , 62.0 , 58.0 , 44.0 , 67.0 ]
])
gt_bboxes7 = np.array([
[ 0 , 28.0 , 31.0 , 55.0 , 63.0 ],
[ 0 , 58.0 , 67.0 , 50.0 , 58.0 ]
])
pred_bboxes_batch.append(pred_bboxes1)
pred_bboxes_batch.append(pred_bboxes2)
pred_bboxes_batch.append(pred_bboxes3)
pred_bboxes_batch.append(pred_bboxes4)
pred_bboxes_batch.append(pred_bboxes5)
pred_bboxes_batch.append(pred_bboxes6)
pred_bboxes_batch.append(pred_bboxes7)
gt_bboxes_batch.append(gt_bboxes1)
gt_bboxes_batch.append(gt_bboxes2)
gt_bboxes_batch.append(gt_bboxes3)
gt_bboxes_batch.append(gt_bboxes4)
gt_bboxes_batch.append(gt_bboxes5)
gt_bboxes_batch.append(gt_bboxes6)
gt_bboxes_batch.append(gt_bboxes7)
class_tp_fp_score_batch = []
for pred_bboxes_per_image, gt_bboxes_per_image in zip(pred_bboxes_batch, gt_bboxes_batch):
class_tp_fp_score = measure_tpfp(pred_bboxes_per_image, gt_bboxes_per_image, 0.5, bbox_format='tlxtlywh')
class_tp_fp_score_batch.append(class_tp_fp_score)
class_tp_fp_score_batch = np.concatenate(class_tp_fp_score_batch, axis=0)
class_tp_fp_score_batch = class_tp_fp_score_batch[np.argsort(class_tp_fp_score_batch[:, 3])[::-1]]
gt_bboxes_batch = np.concatenate(gt_bboxes_batch, axis=0)
print(class_tp_fp_score_batch)
mean_ap = compute_map(class_tp_fp_score_batch, gt_bboxes_batch, num_classes)
print(mean_ap)