-
Notifications
You must be signed in to change notification settings - Fork 51
/
model_pose_ren.py
157 lines (137 loc) · 6.5 KB
/
model_pose_ren.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
import cv2
import numpy as np
import os
import caffe
from caffe.proto import caffe_pb2
import util
class ModelPoseREN(object):
def __init__(self, dataset, center_loader=None,
param=None, use_gpu=False):
self._dataset = dataset
self._center_loader = center_loader
init_proto_name, init_model_name = util.get_model(dataset, 'baseline')
proto_name, model_name = util.get_model(dataset, 'pose_ren')
self._fx, self._fy, self._ux, self._uy = util.get_param(dataset) if param is None else param
self._net = caffe.Net(proto_name, caffe.TEST, weights=model_name)
self._net_init = caffe.Net(init_proto_name, caffe.TEST, weights=init_model_name)
self._input_size = self._net.blobs['data'].shape[-1]
self._cube_size = 150
if use_gpu:
caffe.set_mode_gpu()
caffe.set_device(0)
else:
caffe.set_mode_cpu()
def reset_model(self, dataset, test_id = 0):
self._dataset = dataset
if dataset == 'msra':
init_proto_name, init_model_name = util.get_model(dataset, 'baseline', test_id)
proto_name, model_name = util.get_model(dataset, 'pose_ren', test_id)
else:
init_proto_name, init_model_name = util.get_model(dataset, 'baseline')
proto_name, model_name = util.get_model(dataset, 'pose_ren')
print init_proto_name, init_model_name
print proto_name, model_name
self._net = caffe.Net(proto_name, caffe.TEST, weights=model_name)
self._net_init = caffe.Net(init_proto_name, caffe.TEST, weights=init_model_name)
def detect_images(self, imgs, centers=None):
assert centers is not None or self._center_loader is not None
batch_size = len(imgs)
if centers is None:
centers = np.zeros([batch_size, 3], dtype=np.float32)
for idx, img in enumerate(imgs):
centers[idx, :] = self._center_loader(img)
_, channels, height, width = self._net.blobs['data'].shape
# run Init-CNN
self._net_init.blobs['data'].reshape(batch_size, channels, height, width)
cropped_images = []
for idx in range(batch_size):
cropped_image = self._crop_image(imgs[idx], centers[idx])
cropped_images.append(cropped_image)
self._net_init.blobs['data'].data[idx, ...] = cropped_image
if self._dataset == 'hands17':
init_poses = self._net_init.forward()['predict']
else:
init_poses = self._net_init.forward()['fc3']
# run Pose-REN
prev_pose = init_poses
self._net.blobs['data'].reshape(batch_size, channels, height, width)
_, channels = self._net.blobs['prev_pose'].shape
self._net.blobs['prev_pose'].reshape(batch_size, channels)
for idx in range(batch_size):
self._net.blobs['data'].data[idx, ...] = cropped_images[idx]
for it in xrange(3):
self._net.blobs['prev_pose'].data[...] = prev_pose
if self._dataset == 'hands17':
poses = self._net.forward()['predict']
else:
poses = self._net.forward()['fc3_0']
prev_pose = poses
return self._transform_pose(poses, centers), cropped_images
def detect_image(self, img, center=None):
if center is None:
res, cropped_image = self.detect_images([img])
else:
res, cropped_image = self.detect_images([img], [center])
return res[0, ...], cropped_image[0]
def detect_files(self, base_dir, names, centers=None, dataset=None, max_batch=64, is_flip=False):
assert max_batch > 0
if dataset is None:
dataset = self._dataset
batch_imgs = []
batch_centers = []
results = []
for idx, name in enumerate(names):
img = util.load_image(dataset, os.path.join(base_dir, name),
is_flip=is_flip)
batch_imgs.append(img)
if centers is None:
batch_centers.append(self._center_loader(img))
else:
batch_centers.append(centers[idx, :])
if len(batch_imgs) == max_batch:
res, _= self.detect_images(batch_imgs, batch_centers)
for line in res:
results.append(line)
del batch_imgs[:]
del batch_centers[:]
print('{}/{}'.format(idx + 1, len(names)))
if batch_imgs:
res, _ = self.detect_images(batch_imgs, batch_centers)
for line in res:
results.append(line)
print('done!')
return np.array(results)
def _crop_image(self, img, center, is_debug=False):
xstart = center[0] - self._cube_size / center[2] * self._fx
xend = center[0] + self._cube_size / center[2] * self._fx
ystart = center[1] - self._cube_size / center[2] * self._fy
yend = center[1] + self._cube_size / center[2] * self._fy
src = [(xstart, ystart), (xstart, yend), (xend, ystart)]
dst = [(0, 0), (0, self._input_size - 1), (self._input_size - 1, 0)]
trans = cv2.getAffineTransform(np.array(src, dtype=np.float32),
np.array(dst, dtype=np.float32))
res_img = cv2.warpAffine(img, trans, (self._input_size, self._input_size), None,
cv2.INTER_LINEAR, cv2.BORDER_CONSTANT, center[2] + self._cube_size)
res_img -= center[2]
res_img = np.maximum(res_img, -self._cube_size)
res_img = np.minimum(res_img, self._cube_size)
res_img /= self._cube_size
if is_debug:
img_show = (res_img + 1) / 2;
hehe = cv2.resize(img_show, (512, 512))
cv2.imshow('debug', img_show)
ch = cv2.waitKey(0)
if ch == ord('q'):
exit(0)
return res_img
def _transform_pose(self, poses, centers):
res_poses = np.array(poses) * self._cube_size
num_joint = poses.shape[1] / 3
centers_tile = np.tile(centers, (num_joint, 1, 1)).transpose([1, 0, 2])
res_poses[:, 0::3] = res_poses[:, 0::3] * self._fx / centers_tile[:, :, 2] + centers_tile[:, :, 0]
res_poses[:, 1::3] = res_poses[:, 1::3] * self._fy / centers_tile[:, :, 2] + centers_tile[:, :, 1]
res_poses[:, 2::3] += centers_tile[:, :, 2]
res_poses = np.reshape(res_poses, [poses.shape[0], -1, 3])
if self._dataset == 'nyu':
res_poses = res_poses[:, [6, 7, 8, 9, 10, 11, 12, 13, 3, 4, 5, 1, 2, 0], :]
return res_poses