/
retinaface.py
557 lines (499 loc) · 21.5 KB
/
retinaface.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
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
import sys
import os
import datetime
import time
import numpy as np
import mxnet as mx
from mxnet import ndarray as nd
import cv2
from rcnn.processing.bbox_transform import clip_boxes
from rcnn.processing.generate_anchor import generate_anchors_fpn, anchors_plane
from rcnn.processing.nms import gpu_nms_wrapper, cpu_nms_wrapper
from rcnn.processing.bbox_transform import bbox_overlaps
class RetinaFace:
def __init__(self,
prefix,
epoch,
ctx_id=0,
network='net3',
nms=0.4,
nocrop=False,
decay4=0.5,
vote=False):
self.ctx_id = ctx_id
self.network = network
self.decay4 = decay4
self.nms_threshold = nms
self.vote = vote
self.nocrop = nocrop
self.debug = False
self.fpn_keys = []
self.anchor_cfg = None
pixel_means = [0.0, 0.0, 0.0]
pixel_stds = [1.0, 1.0, 1.0]
pixel_scale = 1.0
self.preprocess = False
_ratio = (1., )
fmc = 3
if fmc == 3:
self._feat_stride_fpn = [32, 16, 8]
self.anchor_cfg = {
'32': {
'SCALES': (32, 16),
'BASE_SIZE': 16,
'RATIOS': _ratio,
'ALLOWED_BORDER': 9999
},
'16': {
'SCALES': (8, 4),
'BASE_SIZE': 16,
'RATIOS': _ratio,
'ALLOWED_BORDER': 9999
},
'8': {
'SCALES': (2, 1),
'BASE_SIZE': 16,
'RATIOS': _ratio,
'ALLOWED_BORDER': 9999
},
}
if self.debug:
print(self._feat_stride_fpn, self.anchor_cfg)
for s in self._feat_stride_fpn:
self.fpn_keys.append('stride%s' % s)
dense_anchor = False
self._anchors_fpn = dict(
zip(
self.fpn_keys,
generate_anchors_fpn(dense_anchor=dense_anchor,
cfg=self.anchor_cfg)))
for k in self._anchors_fpn:
v = self._anchors_fpn[k].astype(np.float32)
self._anchors_fpn[k] = v
self._num_anchors = dict(
zip(self.fpn_keys,
[anchors.shape[0] for anchors in self._anchors_fpn.values()]))
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
if self.ctx_id >= 0:
self.ctx = mx.gpu(self.ctx_id)
self.nms = gpu_nms_wrapper(self.nms_threshold, self.ctx_id)
else:
self.ctx = mx.cpu()
self.nms = cpu_nms_wrapper(self.nms_threshold)
self.pixel_means = np.array(pixel_means, dtype=np.float32)
self.pixel_stds = np.array(pixel_stds, dtype=np.float32)
self.pixel_scale = float(pixel_scale)
if self.debug:
print('means', self.pixel_means)
self.use_landmarks = False
if len(sym) // len(self._feat_stride_fpn) == 3:
self.use_landmarks = True
if self.debug:
print('use_landmarks', self.use_landmarks)
c = len(sym) // len(self._feat_stride_fpn)
sym = sym[(c * 0):]
self._feat_stride_fpn = [32, 16, 8]
print('sym size:', len(sym))
image_size = (640, 640)
self.model = mx.mod.Module(symbol=sym,
context=self.ctx,
label_names=None)
self.model.bind(data_shapes=[('data', (1, 3, image_size[0],
image_size[1]))],
for_training=False)
self.model.set_params(arg_params, aux_params)
def get_input(self, img):
im = img.astype(np.float32)
im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
for i in range(3):
im_tensor[0, i, :, :] = (
im[:, :, 2 - i] / self.pixel_scale -
self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
data = nd.array(im_tensor)
return data
def detect(self, img, threshold=0.5, scales=[1.0], do_flip=False):
if self.debug:
print('in_detect', threshold, scales, do_flip)
proposals_list = []
scores_list = []
landmarks_list = []
timea = datetime.datetime.now()
flips = [0]
if do_flip:
flips = [0, 1]
for im_scale in scales:
for flip in flips:
if im_scale != 1.0:
im = cv2.resize(img,
None,
None,
fx=im_scale,
fy=im_scale,
interpolation=cv2.INTER_LINEAR)
else:
im = img.copy()
if flip:
im = im[:, ::-1, :]
if self.nocrop:
if im.shape[0] % 32 == 0:
h = im.shape[0]
else:
h = (im.shape[0] // 32 + 1) * 32
if im.shape[1] % 32 == 0:
w = im.shape[1]
else:
w = (im.shape[1] // 32 + 1) * 32
_im = np.zeros((h, w, 3), dtype=np.float32)
_im[0:im.shape[0], 0:im.shape[1], :] = im
im = _im
else:
im = im.astype(np.float32)
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('X1 uses', diff.total_seconds(), 'seconds')
im_info = [im.shape[0], im.shape[1]]
im_tensor = np.zeros((1, 3, im.shape[0], im.shape[1]))
for i in range(3):
im_tensor[0, i, :, :] = (
im[:, :, 2 - i] / self.pixel_scale -
self.pixel_means[2 - i]) / self.pixel_stds[2 - i]
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('X2 uses', diff.total_seconds(), 'seconds')
data = nd.array(im_tensor)
db = mx.io.DataBatch(data=(data, ),
provide_data=[('data', data.shape)])
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('X3 uses', diff.total_seconds(), 'seconds')
self.model.forward(db, is_train=False)
net_out = self.model.get_outputs()
for _idx, s in enumerate(self._feat_stride_fpn):
_key = 'stride%s' % s
stride = int(s)
if self.use_landmarks:
idx = _idx * 3
else:
idx = _idx * 2
#print('getting', im_scale, stride, idx, len(net_out), data.shape, file=sys.stderr)
scores = net_out[idx].asnumpy()
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('A uses', diff.total_seconds(), 'seconds')
#print(scores.shape)
#print('scores',stride, scores.shape, file=sys.stderr)
scores = scores[:, self._num_anchors['stride%s' %
s]:, :, :]
idx += 1
bbox_deltas = net_out[idx].asnumpy()
#if DEBUG:
# print 'im_size: ({}, {})'.format(im_info[0], im_info[1])
# print 'scale: {}'.format(im_info[2])
#_height, _width = int(im_info[0] / stride), int(im_info[1] / stride)
height, width = bbox_deltas.shape[2], bbox_deltas.shape[3]
A = self._num_anchors['stride%s' % s]
K = height * width
anchors_fpn = self._anchors_fpn['stride%s' % s]
anchors = anchors_plane(height, width, stride, anchors_fpn)
#print((height, width), (_height, _width), anchors.shape, bbox_deltas.shape, scores.shape, file=sys.stderr)
anchors = anchors.reshape((K * A, 4))
#print('num_anchors', self._num_anchors['stride%s'%s], file=sys.stderr)
#print('HW', (height, width), file=sys.stderr)
#print('anchors_fpn', anchors_fpn.shape, file=sys.stderr)
#print('anchors', anchors.shape, file=sys.stderr)
#print('bbox_deltas', bbox_deltas.shape, file=sys.stderr)
#print('scores', scores.shape, file=sys.stderr)
scores = self._clip_pad(scores, (height, width))
scores = scores.transpose((0, 2, 3, 1)).reshape((-1, 1))
#print('pre', bbox_deltas.shape, height, width)
bbox_deltas = self._clip_pad(bbox_deltas, (height, width))
#print('after', bbox_deltas.shape, height, width)
bbox_deltas = bbox_deltas.transpose((0, 2, 3, 1))
bbox_pred_len = bbox_deltas.shape[3] // A
#print(bbox_deltas.shape)
bbox_deltas = bbox_deltas.reshape((-1, bbox_pred_len))
#print(anchors.shape, bbox_deltas.shape, A, K, file=sys.stderr)
proposals = self.bbox_pred(anchors, bbox_deltas)
proposals = clip_boxes(proposals, im_info[:2])
scores_ravel = scores.ravel()
#print('__shapes', proposals.shape, scores_ravel.shape)
#print('max score', np.max(scores_ravel))
order = np.where(scores_ravel >= threshold)[0]
#scores = scores_ravel[order]
#order = _scores.argsort()[::-1]
#order = order[_order]
proposals = proposals[order, :]
scores = scores[order]
if stride == 4 and self.decay4 < 1.0:
scores *= self.decay4
if flip:
oldx1 = proposals[:, 0].copy()
oldx2 = proposals[:, 2].copy()
proposals[:, 0] = im.shape[1] - oldx2 - 1
proposals[:, 2] = im.shape[1] - oldx1 - 1
proposals[:, 0:4] /= im_scale
proposals_list.append(proposals)
scores_list.append(scores)
if not self.vote and self.use_landmarks:
idx += 1
landmark_deltas = net_out[idx].asnumpy()
landmark_deltas = self._clip_pad(
landmark_deltas, (height, width))
landmark_pred_len = landmark_deltas.shape[1] // A
landmark_deltas = landmark_deltas.transpose(
(0, 2, 3, 1)).reshape(
(-1, 5, landmark_pred_len // 5))
#print(landmark_deltas.shape, landmark_deltas)
landmarks = self.landmark_pred(anchors,
landmark_deltas)
landmarks = landmarks[order, :]
if flip:
landmarks[:, :, 0] = im.shape[1] - landmarks[:, :, 0] - 1
order = [1, 0, 2, 4, 3]
flandmarks = landmarks.copy()
for idx, a in enumerate(order):
flandmarks[:, idx, :] = landmarks[:, a, :]
landmarks = flandmarks
landmarks[:, :, 0:2] /= im_scale
landmarks_list.append(landmarks)
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('B uses', diff.total_seconds(), 'seconds')
proposals = np.vstack(proposals_list)
landmarks = None
if proposals.shape[0] == 0:
if self.use_landmarks:
landmarks = np.zeros((0, 5, 2))
return np.zeros((0, 5)), landmarks
scores = np.vstack(scores_list)
scores_ravel = scores.ravel()
order = scores_ravel.argsort()[::-1]
proposals = proposals[order, :]
scores = scores[order]
if not self.vote and self.use_landmarks:
landmarks = np.vstack(landmarks_list)
landmarks = landmarks[order].astype(np.float32, copy=False)
pre_det = np.hstack((proposals[:, 0:4], scores)).astype(np.float32,
copy=False)
if not self.vote:
keep = self.nms(pre_det)
det = np.hstack((pre_det, proposals[:, 4:]))
det = det[keep, :]
if self.use_landmarks:
landmarks = landmarks[keep]
else:
det = np.hstack((pre_det, proposals[:, 4:]))
det = self.bbox_vote(det)
if self.debug:
timeb = datetime.datetime.now()
diff = timeb - timea
print('C uses', diff.total_seconds(), 'seconds')
return det, landmarks
def detect_center(self, img, threshold=0.5, scales=[1.0], do_flip=False):
det, landmarks = self.detect(img, threshold, scales, do_flip)
if det.shape[0] == 0:
return None, None
bindex = 0
if det.shape[0] > 1:
img_size = np.asarray(img.shape)[0:2]
bounding_box_size = (det[:, 2] - det[:, 0]) * (det[:, 3] -
det[:, 1])
img_center = img_size / 2
offsets = np.vstack([(det[:, 0] + det[:, 2]) / 2 - img_center[1],
(det[:, 1] + det[:, 3]) / 2 - img_center[0]])
offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
bindex = np.argmax(bounding_box_size - offset_dist_squared * 2.0)
bbox = det[bindex, :]
landmark = landmarks[bindex, :, :]
return bbox, landmark
@staticmethod
def check_large_pose(landmark, bbox):
assert landmark.shape == (5, 2)
assert len(bbox) == 4
def get_theta(base, x, y):
vx = x - base
vy = y - base
vx[1] *= -1
vy[1] *= -1
tx = np.arctan2(vx[1], vx[0])
ty = np.arctan2(vy[1], vy[0])
d = ty - tx
d = np.degrees(d)
if d < -180.0:
d += 360.
elif d > 180.0:
d -= 360.0
return d
landmark = landmark.astype(np.float32)
theta1 = get_theta(landmark[0], landmark[3], landmark[2])
theta2 = get_theta(landmark[1], landmark[2], landmark[4])
#print(va, vb, theta2)
theta3 = get_theta(landmark[0], landmark[2], landmark[1])
theta4 = get_theta(landmark[1], landmark[0], landmark[2])
theta5 = get_theta(landmark[3], landmark[4], landmark[2])
theta6 = get_theta(landmark[4], landmark[2], landmark[3])
theta7 = get_theta(landmark[3], landmark[2], landmark[0])
theta8 = get_theta(landmark[4], landmark[1], landmark[2])
#print(theta1, theta2, theta3, theta4, theta5, theta6, theta7, theta8)
left_score = 0.0
right_score = 0.0
up_score = 0.0
down_score = 0.0
if theta1 <= 0.0:
left_score = 10.0
elif theta2 <= 0.0:
right_score = 10.0
else:
left_score = theta2 / theta1
right_score = theta1 / theta2
if theta3 <= 10.0 or theta4 <= 10.0:
up_score = 10.0
else:
up_score = max(theta1 / theta3, theta2 / theta4)
if theta5 <= 10.0 or theta6 <= 10.0:
down_score = 10.0
else:
down_score = max(theta7 / theta5, theta8 / theta6)
mleft = (landmark[0][0] + landmark[3][0]) / 2
mright = (landmark[1][0] + landmark[4][0]) / 2
box_center = ((bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2)
ret = 0
if left_score >= 3.0:
ret = 1
if ret == 0 and left_score >= 2.0:
if mright <= box_center[0]:
ret = 1
if ret == 0 and right_score >= 3.0:
ret = 2
if ret == 0 and right_score >= 2.0:
if mleft >= box_center[0]:
ret = 2
if ret == 0 and up_score >= 2.0:
ret = 3
if ret == 0 and down_score >= 5.0:
ret = 4
return ret, left_score, right_score, up_score, down_score
@staticmethod
def _filter_boxes(boxes, min_size):
""" Remove all boxes with any side smaller than min_size """
ws = boxes[:, 2] - boxes[:, 0] + 1
hs = boxes[:, 3] - boxes[:, 1] + 1
keep = np.where((ws >= min_size) & (hs >= min_size))[0]
return keep
@staticmethod
def _filter_boxes2(boxes, max_size, min_size):
""" Remove all boxes with any side smaller than min_size """
ws = boxes[:, 2] - boxes[:, 0] + 1
hs = boxes[:, 3] - boxes[:, 1] + 1
if max_size > 0:
keep = np.where(np.minimum(ws, hs) < max_size)[0]
elif min_size > 0:
keep = np.where(np.maximum(ws, hs) > min_size)[0]
return keep
@staticmethod
def _clip_pad(tensor, pad_shape):
"""
Clip boxes of the pad area.
:param tensor: [n, c, H, W]
:param pad_shape: [h, w]
:return: [n, c, h, w]
"""
H, W = tensor.shape[2:]
h, w = pad_shape
if h < H or w < W:
tensor = tensor[:, :, :h, :w].copy()
return tensor
@staticmethod
def bbox_pred(boxes, box_deltas):
"""
Transform the set of class-agnostic boxes into class-specific boxes
by applying the predicted offsets (box_deltas)
:param boxes: !important [N 4]
:param box_deltas: [N, 4 * num_classes]
:return: [N 4 * num_classes]
"""
if boxes.shape[0] == 0:
return np.zeros((0, box_deltas.shape[1]))
boxes = boxes.astype(np.float, copy=False)
widths = boxes[:, 2] - boxes[:, 0] + 1.0
heights = boxes[:, 3] - boxes[:, 1] + 1.0
ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
dx = box_deltas[:, 0:1]
dy = box_deltas[:, 1:2]
dw = box_deltas[:, 2:3]
dh = box_deltas[:, 3:4]
pred_ctr_x = dx * widths[:, np.newaxis] + ctr_x[:, np.newaxis]
pred_ctr_y = dy * heights[:, np.newaxis] + ctr_y[:, np.newaxis]
pred_w = np.exp(dw) * widths[:, np.newaxis]
pred_h = np.exp(dh) * heights[:, np.newaxis]
pred_boxes = np.zeros(box_deltas.shape)
# x1
pred_boxes[:, 0:1] = pred_ctr_x - 0.5 * (pred_w - 1.0)
# y1
pred_boxes[:, 1:2] = pred_ctr_y - 0.5 * (pred_h - 1.0)
# x2
pred_boxes[:, 2:3] = pred_ctr_x + 0.5 * (pred_w - 1.0)
# y2
pred_boxes[:, 3:4] = pred_ctr_y + 0.5 * (pred_h - 1.0)
if box_deltas.shape[1] > 4:
pred_boxes[:, 4:] = box_deltas[:, 4:]
return pred_boxes
@staticmethod
def landmark_pred(boxes, landmark_deltas):
if boxes.shape[0] == 0:
return np.zeros((0, landmark_deltas.shape[1]))
boxes = boxes.astype(np.float, copy=False)
widths = boxes[:, 2] - boxes[:, 0] + 1.0
heights = boxes[:, 3] - boxes[:, 1] + 1.0
ctr_x = boxes[:, 0] + 0.5 * (widths - 1.0)
ctr_y = boxes[:, 1] + 0.5 * (heights - 1.0)
pred = landmark_deltas.copy()
for i in range(5):
pred[:, i, 0] = landmark_deltas[:, i, 0] * widths + ctr_x
pred[:, i, 1] = landmark_deltas[:, i, 1] * heights + ctr_y
return pred
def bbox_vote(self, det):
if det.shape[0] == 0:
dets = np.array([[10, 10, 20, 20, 0.002]])
det = np.empty(shape=[0, 5])
while det.shape[0] > 0:
# IOU
area = (det[:, 2] - det[:, 0] + 1) * (det[:, 3] - det[:, 1] + 1)
xx1 = np.maximum(det[0, 0], det[:, 0])
yy1 = np.maximum(det[0, 1], det[:, 1])
xx2 = np.minimum(det[0, 2], det[:, 2])
yy2 = np.minimum(det[0, 3], det[:, 3])
w = np.maximum(0.0, xx2 - xx1 + 1)
h = np.maximum(0.0, yy2 - yy1 + 1)
inter = w * h
o = inter / (area[0] + area[:] - inter)
# nms
merge_index = np.where(o >= self.nms_threshold)[0]
det_accu = det[merge_index, :]
det = np.delete(det, merge_index, 0)
if merge_index.shape[0] <= 1:
if det.shape[0] == 0:
try:
dets = np.row_stack((dets, det_accu))
except:
dets = det_accu
continue
det_accu[:, 0:4] = det_accu[:, 0:4] * np.tile(
det_accu[:, -1:], (1, 4))
max_score = np.max(det_accu[:, 4])
det_accu_sum = np.zeros((1, 5))
det_accu_sum[:, 0:4] = np.sum(det_accu[:, 0:4], axis=0) / np.sum(
det_accu[:, -1:])
det_accu_sum[:, 4] = max_score
try:
dets = np.row_stack((dets, det_accu_sum))
except:
dets = det_accu_sum
dets = dets[0:750, :]
return dets