/
demo_opt.py
477 lines (431 loc) · 18.3 KB
/
demo_opt.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
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# Modified by Peize Sun
# Contact: sunpeize@foxmail.com
# Copyright (c) BaseDetection, Inc. and its affiliates. All Rights Reserved
import json
import os
import numpy as np
import sys
from config import config
PERSON_CLASSES = ['background', 'person']
class Image(object):
def __init__(self, mode):
self.ID = None
self._width = None
self._height = None
self.dtboxes = None
self.gtboxes = None
self.eval_mode = mode
self._ignNum = None
self._gtNum = None
self._dtNum = None
def load(self, record, body_key, head_key, class_names, gtflag):
"""
:meth: read the object from a dict
"""
if "ID" in record and self.ID is None:
self.ID = record['ID']
if "width" in record and self._width is None:
self._width = record["width"]
if "height" in record and self._height is None:
self._height = record["height"]
if gtflag:
self._gtNum = len(record["gtboxes"])
body_bbox, head_bbox = self.load_gt_boxes(record, 'gtboxes', class_names)
if self.eval_mode == 0:
self.gtboxes = body_bbox
self._ignNum = (body_bbox[:, -1] == -1).sum()
elif self.eval_mode == 1:
self.gtboxes = head_bbox
self._ignNum = (head_bbox[:, -1] == -1).sum()
elif self.eval_mode == 2:
gt_tag = np.array(
[body_bbox[i, -1] != -1 and head_bbox[i, -1] != -1
for i in range(len(body_bbox))]
)
self._ignNum = (gt_tag == 0).sum()
self.gtboxes = np.hstack(
(body_bbox[:, :-1], head_bbox[:, :-1], gt_tag.reshape(-1, 1))
)
else:
raise Exception('Unknown evaluation mode!')
if not gtflag:
self._dtNum = len(record["dtboxes"])
if self.eval_mode == 0:
self.dtboxes = self.load_det_boxes(record, 'dtboxes', body_key, 'score')
elif self.eval_mode == 1:
self.dtboxes = self.load_det_boxes(record, 'dtboxes', head_key, 'score')
elif self.eval_mode == 2:
body_dtboxes = self.load_det_boxes(record, 'dtboxes', body_key)
head_dtboxes = self.load_det_boxes(record, 'dtboxes', head_key, 'score')
self.dtboxes = np.hstack((body_dtboxes, head_dtboxes))
else:
raise Exception('Unknown evaluation mode!')
def compare_caltech(self, thres):
"""
:meth: match the detection results with the groundtruth by Caltech matching strategy
:param thres: iou threshold
:type thres: float
:return: a list of tuples (dtbox, imageID), in the descending sort of dtbox.score
"""
if self.dtboxes is None or self.gtboxes is None:
return list()
dtboxes = self.dtboxes if self.dtboxes is not None else list()
gtboxes = self.gtboxes if self.gtboxes is not None else list()
dtboxes = np.array(sorted(dtboxes, key=lambda x: x[-1], reverse=True))
gtboxes = np.array(sorted(gtboxes, key=lambda x: x[-1], reverse=True))
if len(dtboxes):
overlap_iou = self.box_overlap_opr(dtboxes, gtboxes[gtboxes[:, -1] > 0], True)
overlap_ioa = self.box_overlap_opr(dtboxes, gtboxes[gtboxes[:, -1] <= 0], False)
ign = np.any(overlap_ioa > thres, 1)
pos = np.any(overlap_iou > thres, 1)
else:
return list()
scorelist = list()
for i, dt in enumerate(dtboxes):
maxpos = np.argmax(overlap_iou[i])
if overlap_iou[i, maxpos] > thres:
overlap_iou[:, maxpos] = 0
scorelist.append((dt, 1, self.ID, pos[i]))
elif not ign[i]:
scorelist.append((dt, 0, self.ID, pos[i]))
return scorelist
def compare_caltech_union(self, thres):
"""
:meth: match the detection results with the groundtruth by Caltech matching strategy
:param thres: iou threshold
:type thres: float
:return: a list of tuples (dtbox, imageID), in the descending sort of dtbox.score
"""
dtboxes = self.dtboxes if self.dtboxes is not None else list()
gtboxes = self.gtboxes if self.gtboxes is not None else list()
if len(dtboxes) == 0:
return list()
dt_matched = np.zeros(dtboxes.shape[0])
gt_matched = np.zeros(gtboxes.shape[0])
dtboxes = np.array(sorted(dtboxes, key=lambda x: x[-1], reverse=True))
gtboxes = np.array(sorted(gtboxes, key=lambda x: x[-1], reverse=True))
dt_body_boxes = np.hstack((dtboxes[:, :4], dtboxes[:, -1][:, None]))
dt_head_boxes = dtboxes[:, 4:8]
gt_body_boxes = np.hstack((gtboxes[:, :4], gtboxes[:, -1][:, None]))
gt_head_boxes = gtboxes[:, 4:8]
overlap_iou = self.box_overlap_opr(dt_body_boxes, gt_body_boxes, True)
overlap_head = self.box_overlap_opr(dt_head_boxes, gt_head_boxes, True)
overlap_ioa = self.box_overlap_opr(dt_body_boxes, gt_body_boxes, False)
scorelist = list()
for i, dt in enumerate(dtboxes):
maxpos = -1
maxiou = thres
for j, gt in enumerate(gtboxes):
if gt_matched[j] == 1:
continue
if gt[-1] > 0:
o_body = overlap_iou[i][j]
o_head = overlap_head[i][j]
if o_body > maxiou and o_head > maxiou:
maxiou = o_body
maxpos = j
else:
if maxpos >= 0:
break
else:
o_body = overlap_ioa[i][j]
if o_body > thres:
maxiou = o_body
maxpos = j
if maxpos >= 0:
if gtboxes[maxpos, -1] > 0:
gt_matched[maxpos] = 1
dt_matched[i] = 1
scorelist.append((dt, 1, self.ID))
else:
dt_matched[i] = -1
else:
dt_matched[i] = 0
scorelist.append((dt, 0, self.ID))
return scorelist
def box_overlap_opr(self, dboxes: np.ndarray, gboxes: np.ndarray, if_iou) -> np.ndarray:
eps = 1e-6
assert dboxes.shape[-1] >= 4 and gboxes.shape[-1] >= 4
N, K = dboxes.shape[0], gboxes.shape[0]
dtboxes = np.tile(np.expand_dims(dboxes, axis=1), (1, K, 1))
gtboxes = np.tile(np.expand_dims(gboxes, axis=0), (N, 1, 1))
iw = (np.minimum(dtboxes[:, :, 2], gtboxes[:, :, 2])
- np.maximum(dtboxes[:, :, 0], gtboxes[:, :, 0]))
ih = (np.minimum(dtboxes[:, :, 3], gtboxes[:, :, 3])
- np.maximum(dtboxes[:, :, 1], gtboxes[:, :, 1]))
inter = np.maximum(0, iw) * np.maximum(0, ih)
dtarea = (dtboxes[:, :, 2] - dtboxes[:, :, 0]) * (dtboxes[:, :, 3] - dtboxes[:, :, 1])
if if_iou:
gtarea = (gtboxes[:, :, 2] - gtboxes[:, :, 0]) * (gtboxes[:, :, 3] - gtboxes[:, :, 1])
ious = inter / (dtarea + gtarea - inter + eps)
else:
ious = inter / (dtarea + eps)
return ious
def clip_all_boader(self):
def _clip_boundary(boxes, height, width):
assert boxes.shape[-1] >= 4
boxes[:, 0] = np.minimum(np.maximum(boxes[:, 0], 0), width - 1)
boxes[:, 1] = np.minimum(np.maximum(boxes[:, 1], 0), height - 1)
boxes[:, 2] = np.maximum(np.minimum(boxes[:, 2], width), 0)
boxes[:, 3] = np.maximum(np.minimum(boxes[:, 3], height), 0)
return boxes
assert self.dtboxes.shape[-1] >= 4
assert self.gtboxes.shape[-1] >= 4
assert self._width is not None and self._height is not None
if self.eval_mode == 2:
self.dtboxes[:, :4] = _clip_boundary(self.dtboxes[:, :4], self._height, self._width)
self.gtboxes[:, :4] = _clip_boundary(self.gtboxes[:, :4], self._height, self._width)
self.dtboxes[:, 4:8] = _clip_boundary(self.dtboxes[:, 4:8], self._height, self._width)
self.gtboxes[:, 4:8] = _clip_boundary(self.gtboxes[:, 4:8], self._height, self._width)
else:
self.dtboxes = _clip_boundary(self.dtboxes, self._height, self._width)
self.gtboxes = _clip_boundary(self.gtboxes, self._height, self._width)
def load_gt_boxes(self, dict_input, key_name, class_names):
assert key_name in dict_input
if len(dict_input[key_name]) < 1:
return np.empty([0, 5])
head_bbox = []
body_bbox = []
for rb in dict_input[key_name]:
if rb['tag'] in class_names:
body_tag = class_names.index(rb['tag'])
head_tag = 1
else:
body_tag = -1
head_tag = -1
if 'extra' in rb:
if 'ignore' in rb['extra']:
if rb['extra']['ignore'] != 0:
body_tag = -1
head_tag = -1
if 'head_attr' in rb:
if 'ignore' in rb['head_attr']:
if rb['head_attr']['ignore'] != 0:
head_tag = -1
# head_bbox.append(np.hstack((rb['hbox'], head_tag)))
body_bbox.append((*rb['fbox'], body_tag))
# head_bbox = np.array(head_bbox)
# head_bbox[:, 2:4] += head_bbox[:, :2]
body_bbox = np.array(body_bbox)
body_bbox[:, 2:4] += body_bbox[:, :2]
return body_bbox, head_bbox
def load_det_boxes(self, dict_input, key_name, key_box, key_score=None, key_tag=None):
assert key_name in dict_input
if len(dict_input[key_name]) < 1:
return np.empty([0, 5])
else:
assert key_box in dict_input[key_name][0]
if key_score:
assert key_score in dict_input[key_name][0]
if key_tag:
assert key_tag in dict_input[key_name][0]
if key_score:
if key_tag:
bboxes = np.vstack(
[
np.hstack(
(rb[key_box], rb[key_score], rb[key_tag])
) for rb in dict_input[key_name]
]
)
else:
bboxes = np.array([(*rb[key_box], rb[key_score]) for rb in dict_input[key_name]])
else:
if key_tag:
bboxes = np.vstack(
[np.hstack((rb[key_box], rb[key_tag])) for rb in dict_input[key_name]]
)
else:
bboxes = np.vstack([rb[key_box] for rb in dict_input[key_name]])
bboxes[:, 2:4] += bboxes[:, :2]
return bboxes
def compare_voc(self, thres):
"""
:meth: match the detection results with the groundtruth by VOC matching strategy
:param thres: iou threshold
:type thres: float
:return: a list of tuples (dtbox, imageID), in the descending sort of dtbox.score
"""
if self.dtboxes is None:
return list()
dtboxes = self.dtboxes
gtboxes = self.gtboxes if self.gtboxes is not None else list()
dtboxes.sort(key=lambda x: x.score, reverse=True)
gtboxes.sort(key=lambda x: x.ign)
scorelist = list()
for i, dt in enumerate(dtboxes):
maxpos = -1
maxiou = thres
for j, gt in enumerate(gtboxes):
overlap = dt.iou(gt)
if overlap > maxiou:
maxiou = overlap
maxpos = j
if maxpos >= 0:
if gtboxes[maxpos].ign == 0:
gtboxes[maxpos].matched = 1
dtboxes[i].matched = 1
scorelist.append((dt, self.ID))
else:
dtboxes[i].matched = -1
else:
dtboxes[i].matched = 0
scorelist.append((dt, self.ID))
return scorelist
class Database(object):
def __init__(self, gtpath=None, dtpath=None, body_key=None, head_key=None, mode=0):
"""
mode=0: only body; mode=1: only head
"""
self.images = dict()
self.eval_mode = mode
self.loadData(gtpath, body_key, head_key, if_gt=True)
self.loadData(dtpath, body_key, head_key, if_gt=False)
self._ignNum = sum([self.images[i]._ignNum for i in self.images])
self._gtNum = sum([self.images[i]._gtNum for i in self.images])
self._imageNum = len(self.images)
self.scorelist = None
def loadData(self, fpath, body_key=None, head_key=None, if_gt=True):
assert os.path.isfile(fpath), fpath + " does not exist!"
with open(fpath, "r") as f:
lines = f.readlines()
records = [json.loads(line.strip('\n')) for line in lines]
if if_gt:
for record in records:
self.images[record["ID"]] = Image(self.eval_mode)
self.images[record["ID"]].load(record, body_key, head_key, PERSON_CLASSES, True)
else:
for record in records:
self.images[record["ID"]].load(record, body_key, head_key, PERSON_CLASSES, False)
self.images[record["ID"]].clip_all_boader()
def compare(self, thres=0.5, matching=None):
"""
match the detection results with the groundtruth in the whole database
"""
assert matching is None or matching == "VOC", matching
scorelist = list()
for ID in self.images:
if matching == "VOC":
result = self.images[ID].compare_voc(thres)
else:
result = self.images[ID].compare_caltech(thres)
scorelist.extend(result)
# In the descending sort of dtbox score.
scorelist.sort(key=lambda x: x[0][-1], reverse=True)
self.scorelist = scorelist
def eval_MR(self, ref="CALTECH_-2", fppiX=None, fppiY=None):
"""
evaluate by Caltech-style log-average miss rate
ref: str - "CALTECH_-2"/"CALTECH_-4"
"""
# find greater_than
def _find_gt(lst, target):
for idx, item in enumerate(lst):
if item >= target:
return idx
return len(lst) - 1
assert ref == "CALTECH_-2" or ref == "CALTECH_-4", ref
if ref == "CALTECH_-2":
# CALTECH_MRREF_2: anchor points (from 10^-2 to 1) as in P.Dollar's paper
ref = [0.0100, 0.0178, 0.03160, 0.0562, 0.1000, 0.1778, 0.3162, 0.5623, 1.000]
else:
# CALTECH_MRREF_4: anchor points (from 10^-4 to 1) as in S.Zhang's paper
ref = [0.0001, 0.0003, 0.00100, 0.0032, 0.0100, 0.0316, 0.1000, 0.3162, 1.000]
if self.scorelist is None:
self.compare()
tp, fp = 0.0, 0.0
if fppiX is None or fppiY is None:
fppiX, fppiY = list(), list()
for i, item in enumerate(self.scorelist):
if item[1] == 1:
tp += 1.0
elif item[1] == 0:
fp += 1.0
fn = (self._gtNum - self._ignNum) - tp
recall = tp / (tp + fn)
missrate = 1.0 - recall
fppi = fp / self._imageNum
fppiX.append(fppi)
fppiY.append(missrate)
score = list()
for pos in ref:
argmin = _find_gt(fppiX, pos)
if argmin >= 0:
score.append(fppiY[argmin])
score = np.array(score)
MR = np.exp(np.log(score).mean())
return MR, (fppiX, fppiY)
def eval_AP(self):
"""
:meth: evaluate by average precision
"""
# calculate general ap score
def _calculate_map(recall, precision):
assert len(recall) == len(precision)
area = 0
for i in range(1, len(recall)):
delta_h = (precision[i - 1] + precision[i]) / 2
delta_w = recall[i] - recall[i - 1]
area += delta_w * delta_h
return area
tp, fp, dp = 0.0, 0.0, 0.0
rpX, rpY = list(), list()
total_gt = self._gtNum - self._ignNum
total_images = self._imageNum
fpn = []
dpn = []
recalln = []
thr = []
fppi = []
mr = []
for i, item in enumerate(self.scorelist):
if item[1] == 1:
tp += 1.0
elif item[1] == 0:
fp += 1.0
dp += item[-1]
fn = total_gt - tp
recall = tp / (tp + fn)
precision = tp / (tp + fp)
rpX.append(recall)
rpY.append(precision)
fpn.append(fp)
dpn.append(dp)
recalln.append(tp)
thr.append(item[0][-1])
fppi.append(fp / total_images)
mr.append(1 - recall)
AP = _calculate_map(rpX, rpY)
return AP, recall, (rpX, rpY, thr, fpn, dpn, recalln, fppi, mr)
def _evaluate_predictions_on_crowdhuman(gt_path, dt_path, target_key="box", mode=0):
"""
Evaluate the coco results using COCOEval API.
"""
database = Database(gt_path, dt_path, target_key, None, mode)
database.compare()
AP, recall, data = database.eval_AP()
mMR, _ = database.eval_MR(fppiX=data[-2], fppiY=data[-1])
return AP, mMR, computeJaccard(gt_path, dt_path), recall
def computeJaccard(gt_path, dt_path):
from demo import load_func, common_process, worker
dt = load_func(dt_path)
gt = load_func(gt_path)
ji = 0.
for i in range(1, 10):
results = common_process(worker, dt, 4, gt, i * 0.1, 0.5)
ji = max(ji, np.sum([rb['ratio'] for rb in results]) / 4370)
return ji
if __name__ == "__main__":
gt_path = '/data/CrowdHuman/odformat/crowdhuman_val.odgt'
save_file = f'record_{sys.argv[1]}.txt'
fname = os.path.join(sys.argv[1], 'epoch-{}.human'.format(sys.argv[2]))
fpath = os.path.join(config.eval_dir, fname)
eval_results = _evaluate_predictions_on_crowdhuman(gt_path, fpath)
metric_names = ["AP", "MR", "JI", "Recall"]
for k, v in zip(metric_names, eval_results):
print(f"{k}: {v}")
with open(save_file, "a") as f:
results = [f"{k}: {v:.4f}" for k, v in zip(metric_names, eval_results)]
f.write(", ".join(results) + "\n")