forked from akanazawa/hmr
-
-
Notifications
You must be signed in to change notification settings - Fork 5
/
generate_3DMSMT_bg2.py
277 lines (248 loc) · 9.85 KB
/
generate_3DMSMT_bg2.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
"""
Demo of HMR.
Note that HMR requires the bounding box of the person in the image. The best performance is obtained when max length of the person in the image is roughly 150px.
When only the image path is supplied, it assumes that the image is centered on a person whose length is roughly 150px.
Alternatively, you can supply output of the openpose to figure out the bbox and the right scale factor.
Sample usage:
# On images on a tightly cropped image around the person
python -m generate_3DMarket --market_path ../Market/pytorch/
python -m generate_3DMarket --market_path ../Duke/pytorch/
# On images, with openpose output
python -m demo --img_path data/random.jpg --json_path data/random_keypoints.json
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
from absl import flags
import numpy as np
import skimage.io as io
from skimage.transform import resize
import tensorflow as tf
import os
from PIL import Image
from src.util import renderer as vis_util
from src.util import image as img_util
from src.tf_smpl import projection as proj_util
from src.util import openpose as op_util
import src.config
from src.RunModel import RunModel
flags.DEFINE_string('market_path', '../MSMT/pytorch/', 'Image to run')
flags.DEFINE_string(
'json_path', None,
'If specified, uses the openpose output to crop the image.')
def visualize(img, proc_param, joints, verts, cam):
"""
Renders the result in original image coordinate frame.
"""
cam_for_render, vert_shifted, joints_orig = vis_util.get_original(
proc_param, verts, cam, joints, img_size=img.shape[:2])
#visualize(img, proc_param, joints[0], verts[0], cams[0])
# Render results
skel_img = vis_util.draw_skeleton(img, joints_orig)
rend_img_overlay = renderer(
#vert_shifted, cam=None, img=img, do_alpha=True)
vert_shifted, cam=cam_for_render, img=img, do_alpha=True)
rend_img = renderer(
vert_shifted, cam=cam_for_render, img_size=img.shape[:2])
rend_img_vp1 = renderer.rotated(
vert_shifted, 60, cam=cam_for_render, img_size=img.shape[:2])
rend_img_vp2 = renderer.rotated(
vert_shifted, -60, cam=cam_for_render, img_size=img.shape[:2])
import matplotlib
matplotlib.use('agg')
import matplotlib.pyplot as plt
# plt.ion()
fig = plt.figure()
plt.figure(1)
plt.clf()
plt.subplot(231)
plt.imshow(img)
plt.title('input')
plt.axis('off')
plt.subplot(232)
plt.imshow(skel_img)
plt.title('joint projection')
plt.axis('off')
plt.subplot(233)
plt.imshow(rend_img_overlay)
plt.title('3D Mesh overlay')
plt.axis('off')
plt.subplot(234)
plt.imshow(rend_img)
result = Image.fromarray(rend_img)
result.save('mesh.jpg')
plt.title('3D mesh')
plt.axis('off')
plt.subplot(235)
plt.imshow(rend_img_vp1)
plt.title('diff vp')
plt.axis('off')
plt.subplot(236)
plt.imshow(rend_img_vp2)
plt.title('diff vp')
plt.axis('off')
plt.draw()
fig.savefig('demo.jpg')
# import ipdb
# ipdb.set_trace()
def preprocess_image(img_path, json_path=None):
#img = io.imread(img_path)
#img = Image.fromarray(img)
img = Image.open(img_path)
img = img.resize((128,256))
img = np.array(img)
#img = resize(img, (128 , 64))
if img.shape[2] == 4:
img = img[:, :, :3]
if json_path is None:
if np.max(img.shape[:2]) != config.img_size:
print('Resizing so the max image size is %d..' % config.img_size)
scale = (float(config.img_size) / np.max(img.shape[:2]))
else:
scale = 1.
center = np.round(np.array(img.shape[:2]) / 2).astype(int)
# image center in (x,y)
center = center[::-1]
else:
scale, center = op_util.get_bbox(json_path)
crop, proc_param = img_util.scale_and_crop(img, scale, center,
config.img_size)
# Normalize image to [-1, 1]
crop = 2 * ((crop / 255.) - 0.5)
return crop, proc_param, img
def main(dir_path, json_path=None):
if not os.path.exists('../3D-Person-reID/3DMSMT+bg2'):
os.mkdir('../3D-Person-reID/3DMSMT+bg2')
sess = tf.Session()
model = RunModel(config, sess=sess)
#for split in ['train', 'train_all', 'val', 'gallery', 'query']:
for split in ['val', 'train']:
for root, dirs, files in os.walk(dir_path+split, topdown=True):
for img_path in files:
if not img_path[-3:]=='jpg':
continue
img_path = root +'/' + img_path
print(img_path)
input_img, proc_param, img = preprocess_image(img_path, json_path)
input_img = np.expand_dims(input_img, 0)
# Theta is the 85D vector holding [camera, pose, shape]
# where camera is 3D [s, tx, ty]
# pose is 72D vector holding the rotation of 24 joints of SMPL in axis angle format
# shape is 10D shape coefficients of SMPL
joints, verts, cams, joints3d, theta = model.predict(
input_img, get_theta=True)
# scaling and translation
save_mesh(img, img_path, split, proc_param, joints[0], verts[0], cams[0])
def save_mesh(img, img_path, split, proc_param, joints, verts, cam):
cam_for_render, vert_3d, joints_orig = vis_util.get_original(
proc_param, verts, cam, joints, img_size=img.shape[:2])
cam_for_render, vert_shifted = cam, verts
#print(proc_param)
#print(vert_shifted)
camera = np.reshape(cam_for_render, [1,3])
w, h, _ = img.shape
imgsize = max(w,h)
# project to 2D
vert_2d = verts[:, :2] + camera[:, 1:]
vert_2d = vert_2d * camera[0,0]
img_copy = img.copy()
face_path = './src/tf_smpl/smpl_faces.npy'
faces = np.load(face_path)
obj_mesh_name = '../3D-Person-reID/3DMSMT+bg2/%s/%s/%s.obj'%( split, os.path.basename(os.path.dirname(img_path)), os.path.basename(img_path) )
store_dir = os.path.dirname(obj_mesh_name)
if not os.path.exists(os.path.dirname(store_dir)):
os.mkdir(os.path.dirname(store_dir))
if not os.path.exists(store_dir):
os.mkdir(store_dir)
foreground_index_2d = np.zeros((w,h))+99999
foreground_value_2d = np.zeros((w,h))+99999
background = np.zeros((w,h))
index = 6891
with open(obj_mesh_name, 'w') as fp:
w, h, _ = img.shape
imgsize = max(w,h)
# Decide Forground
for i in range(vert_2d.shape[0]):
v2 = vert_2d[i,:]
v3 = vert_3d[i,:]
z = v3[2]
x = int(round( (v2[1]+1)*0.5*imgsize ))
y = int(round( (v2[0]+1)*0.5*imgsize ))
if w<h:
x = int(round(x -h/2 + w/2))
else:
y = int(round(y - w/2 + h/2))
x = max(0, min(x, w-1))
y = max(0, min(y, h-1))
if z < foreground_value_2d[x,y]:
foreground_index_2d[x,y] = i
foreground_value_2d[x,y] = z
#s smooth
z_max = max(vert_3d[:, 2])- min(vert_3d[:, 2])
for t in range(10):
for i in range(1,w-1):
for j in range(1,h-1):
center= foreground_value_2d[i,j]
if foreground_index_2d[i-1,j] != 999999 and foreground_value_2d[i-1,j]>center+0.05:
foreground_index_2d[i-1,j] = 999999
foreground_value_2d[i-1,j] = 999999
if foreground_index_2d[i,j-1] != 999999 and foreground_value_2d[i,j-1]>center+0.05:
foreground_index_2d[i,j-1] = 999999
foreground_value_2d[i,j-1] = 999999
# Draw Color
for i in range(vert_2d.shape[0]):
v2 = vert_2d[i,:]
v3 = verts[i,:]
z = v3[2]
x = int(round( (v2[1]+1)*0.5*imgsize ))
y = int(round( (v2[0]+1)*0.5*imgsize ))
if w<h:
x = int(round(x -h/2 + w/2))
else:
y = int(round(y - w/2 + h/2))
x = max(0, min(x, w-1))
y = max(0, min(y, h-1))
if i == foreground_index_2d[x,y]:
c = img[x, y, :]/255.0
img_copy[x,y,:] = 0
else:
c = [1,1,1]
continue
fp.write( 'v %f %f %f %f %f %f\n' % ( v3[0], v3[1], v3[2], c[0], c[1], c[2]) )
# 2D to 3D mapping
for i in range(w):
for j in range(h):
vx, vy = i, j
if foreground_index_2d[i,j] < 99999:
continue
if w<h:
vx = vx + h/2 - w/2
else:
vy = vy + w/2 - h/2
vx = vx/imgsize *2 - 1
vy = vy/imgsize *2 - 1
vy /= camera[0,0]
vy -= camera[:, 1]
vx /= camera[0,0]
vx -= camera[:, 2]
vz = np.mean(verts[:,2])
c = img[i,j,:]/255.0
fp.write( 'v %f %f %f %f %f %f\n' % ( vy, vx, vz, c[0], c[1], c[2]) )
background[i,j] = index
index +=1
for f in faces: # Faces are 1-based, not 0-based in obj files
fp.write( 'f %d %d %d\n' % (f[0] + 1, f[1] + 1, f[2] + 1) )
break # skip for small file
#count = 0
#for i in range(1,w):
# for j in range(1,h):
# fp.write( 'f %d %d %d %d\n' % (background[i,j], background[i-1,j] ,background[i,j-1] , background[i-1, j-1]))
if __name__ == '__main__':
config = flags.FLAGS
config(sys.argv)
# Using pre-trained model, change this to use your own.
config.load_path = src.config.PRETRAINED_MODEL
config.batch_size = 1
renderer = vis_util.SMPLRenderer(face_path=config.smpl_face_path)
main(config.market_path, config.json_path)