-
Notifications
You must be signed in to change notification settings - Fork 5
/
util.py
319 lines (248 loc) · 10.4 KB
/
util.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
import torch
import pdb
import torch.nn.functional as F
import numpy as np
import transformations
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def make_warped_stack(mask, src_in, trans_in):
for i in range(mask.shape[1]):
mask_i = mask[:, i,:, :].unsqueeze(1).repeat(1,3,1,1)
src_masked = mask_i * src_in
if i == 0:
warps = src_masked
else:
warp_i = affine_warp(src_masked, trans_in[:, :, :, i-1])
warps = torch.cat([warps, warp_i], 1)
return warps
def affine_warp(im, theta):
num_batch = im.shape[0]
height = im.shape[2]
width = im.shape[3]
y_t, x_t = torch.meshgrid(torch.arange(0,height), torch.arange(0,width))
y_t = y_t.float()
x_t = x_t.float()
x_t_flat = x_t.reshape(1,-1)
y_t_flat = y_t.reshape(1,-1)
ones = torch.ones_like(x_t_flat)
# print(x_t_flat.device,y_t_flat.device,ones.device)
grid = torch.cat((x_t_flat, y_t_flat, ones),dim = 0).to(device)
grid = grid.unsqueeze(0)
grid = grid.repeat(num_batch,1,1)
T_g = torch.matmul(theta, grid)
x_s = T_g[:,0,:].reshape(num_batch, height, width)
y_s = T_g[:,1,:].reshape(num_batch, height, width)
return interpolate(im, x_s, y_s)
def repeat(x,n_repeats):
rep = torch.ones(1,n_repeats).float()
x = torch.matmul(x.reshape(-1,1).float(),rep)
return x.flatten()
def interpolate(im,x,y):
im = F.pad(im,[1,1,1,1],"reflect")
num_batch = im.shape[0]
height = im.shape[2]
width = im.shape[3]
channels = im.shape[1]
out_height = x.shape[1]
out_width = x.shape[2]
x = x.flatten()
y = x.flatten()
x = x+1
y = y+1
max_x = width - 1
max_y = height - 1
x0 = torch.floor(x)
x1 = x0 + 1
y0 = torch.floor(y)
y1 = y0 + 1
x0 = x0.clamp(0, max_x)
x1 = x1.clamp(0, max_x)
y0 = y0.clamp(0, max_y)
y1 = y1.clamp(0, max_y)
base = repeat(torch.arange(num_batch)*width*height, out_height*out_width).to(device)
base_y0 = base + y0*width
base_y1 = base + y1*width
idx_a = base_y0 + x0
idx_b = base_y1 + x0
idx_c = base_y0 + x1
idx_d = base_y1 + x1
# use indices to lookup pixels in the flat image and restore
# channels dim
im_flat = im.permute(0,2,3,1).reshape(-1,channels)
# im_flat = tf.cast(im_flat, 'float32')
Ia = im_flat[idx_a.long()]
Ib = im_flat[idx_b.long()]
Ic = im_flat[idx_c.long()]
Id = im_flat[idx_d.long()]
# and finally calculate interpolated values
x1_f = x1.float()
y1_f = y1.float()
dx = x1_f - x
dy = y1_f - y
wa = torch.unsqueeze(dx * dy, 1)
wb = torch.unsqueeze(dx * (1-dy), 1)
wc = torch.unsqueeze((1-dx) * dy, 1)
wd = torch.unsqueeze((1-dx) * (1-dy), 1)
output = wa*Ia + wb*Ib + wc*Ic + wd*Id
output = output.reshape(-1,out_height,out_width,channels).permute(0,3,1,2)
return output
def fast_gaussian_torch(img_width, img_height, center, var_x, var_y, L=49):
tmp = torch.zeros((img_height,img_width)).to(device)
if var_x<10 and var_y<10:
L=11
patch = make_gaussian_map_torch(L, L, (L//2, L//2), var_x, var_y, 0)
xx = int(center[0])
yy = int(center[1])
start_H = yy-patch.shape[0]//2
end_H = clamp_min_max(yy-patch.shape[0]//2+patch.shape[0], min=0, max=img_height-1)
start_W = xx-patch.shape[1]//2
end_W = clamp_min_max(xx-patch.shape[1]//2+patch.shape[1], min=0, max=img_width-1)
flag = 4
if start_H <0 and start_W<0:
flag = 0
if start_H <0 and start_W>0:
flag = 1
if start_H >0 and start_W<0:
flag = 2
if start_H >0 and start_W>0:
flag = 3
start_H = clamp_min_max(start_H, min=0, max=img_height)
start_W = clamp_min_max(xx-patch.shape[1]//2, min=0, max=img_width)
patch = crop_patch(patch, end_H-start_H, end_W-start_W, flag)
if tmp[start_H: end_H , start_W: end_W].shape == patch.shape:
tmp[start_H: end_H , start_W: end_W] = patch
return tmp
def make_gaussian_map(img_width, img_height, center, var_x, var_y, theta):
xv, yv = np.meshgrid(np.array(range(img_width)), np.array(range(img_height)),
sparse=False, indexing='xy')
a = np.cos(theta) ** 2 / (2 * var_x) + np.sin(theta) ** 2 / (2 * var_y)
b = -np.sin(2 * theta) / (4 * var_x) + np.sin(2 * theta) / (4 * var_y)
c = np.sin(theta) ** 2 / (2 * var_x) + np.cos(theta) ** 2 / (2 * var_y)
return np.exp(-(a * (xv - center[0]) * (xv - center[0]) +
2 * b * (xv - center[0]) * (yv - center[1]) +
c * (yv - center[1]) * (yv - center[1])))
def make_gaussian_map_torch(img_width, img_height, center, var_x, var_y, theta):
yv, xv = torch.meshgrid(torch.arange(0,img_height), torch.arange(0,img_width))
yv = yv.to(device)
xv = xv.to(device)
a = np.cos(theta) ** 2 / (2 * var_x) + np.sin(theta) ** 2 / (2 * var_y)
b = -np.sin(2 * theta) / (4 * var_x) + np.sin(2 * theta) / (4 * var_y)
c = np.sin(theta) ** 2 / (2 * var_x) + np.cos(theta) ** 2 / (2 * var_y)
return torch.exp(-(a * (xv - center[0]) * (xv - center[0]) +
2 * b * (xv - center[0]) * (yv - center[1]) +
c * (yv - center[1]) * (yv - center[1])))
def make_limb_masks(limbs, joints, img_width, img_height, sigma_perp_root):
n_limbs = len(limbs)
mask = torch.zeros((img_height, img_width, n_limbs)).to(device)
# Gaussian sigma perpendicular to the limb axis.
sigma_perp = np.array(sigma_perp_root) ** 2
for i in range(n_limbs):
n_joints_for_limb = len(limbs[i])
p = np.zeros((n_joints_for_limb, 2))
for j in range(n_joints_for_limb):
p[j, :] = [joints[limbs[i][j], 0], joints[limbs[i][j], 1]]
if n_joints_for_limb == 4:
p_top = np.mean(p[0:2, :], axis=0)
p_bot = np.mean(p[2:4, :], axis=0)
p = np.vstack((p_top, p_bot))
center = np.mean(p, axis=0)
sigma_parallel = np.max([5, (np.sum((p[1, :] - p[0, :]) ** 2)) / 1.5])
theta = np.arctan2(p[1, 1] - p[0, 1], p[0, 0] - p[1, 0])
mask_i = make_gaussian_map_torch(img_width, img_height, center, sigma_parallel, sigma_perp[i], theta)
mask[:, :, i] = mask_i / (torch.max(mask_i) + 1e-6)
return mask.permute(2,0,1)
def make_cluster_kp(cluster, joints, img_width, img_height, var_root):
n_cluster = len(cluster)
pose = torch.zeros((img_height, img_width, n_cluster)).to(device)
var = np.array(var_root)**2
for i in range(n_cluster):
n_joints_for_cluster = len(cluster[i])
kp_canvas = torch.zeros((img_height,img_width)).to(device)
for j in range(n_joints_for_cluster):
tmp =fast_gaussian_torch(img_width,img_height,joints[cluster[i][j]],var[i],var[i])
kp_canvas += tmp
pose[:,:,i] = kp_canvas
return pose.permute(2,0,1)
def get_limb_transforms(limbs, joints1, joints2):
n_limbs = len(limbs)
Ms = np.zeros((2, 3, n_limbs))
for i in range(n_limbs):
n_joints_for_limb = len(limbs[i])
p1 = np.zeros((n_joints_for_limb, 2))
p2 = np.zeros((n_joints_for_limb, 2))
for j in range(n_joints_for_limb):
p1[j, :] = [joints1[limbs[i][j], 0], joints1[limbs[i][j], 1]]
p2[j, :] = [joints2[limbs[i][j], 0], joints2[limbs[i][j], 1]]
tform = transformations.make_similarity(p2, p1, False)
Ms[:, :, i] = np.array([[tform[1], -tform[3], tform[0]], [tform[3], tform[1], tform[2]]])
return Ms
def batchify_mask_prior(joints, img_width, img_height, sigma_perp_root = [35, 25, 25, 20, 25, 20, 10, 10]):
limbs = [[0,8,9],[1,2,5],[2,3],[3,4],[5,6],[6,7],range(101,122),range(80,101)]
batch_size = joints.shape[0]
result = []
for i in range(batch_size):
limb_masks = make_limb_masks(limbs,joints[i],img_width,img_height, sigma_perp_root)
bg_mask = (1.0 - torch.max(limb_masks, dim=0)[0]).unsqueeze(0)
mask_prior = torch.log(torch.cat((bg_mask, limb_masks), dim=0) + 1e-10)
result.append(mask_prior)
return torch.stack(result, dim=0)
def batchify_cluster_kp(joints, img_width, img_height, var_root = [5,5,5,5,5,5,5,2,1,1,1,1,1]):
cluster = [[1],[2],[3],[4],[5],[6],[7],range(10,27),[8,9,78,79]+list(range(28,36))+list(range(47,57)),range(37,46),range(58,78),range(101,122),range(80,101)]
batch_size = joints.shape[0]
result = []
for i in range(batch_size):
result.append(make_cluster_kp(cluster,joints[i],img_width,img_height, var_root))
return torch.stack(result, dim=0)
def clamp_min_max(x, min=0, max=99999):
if x<min:
return min
elif x>max:
return max
return x
def crop_patch(patch, valid_H, valid_W, type):
H, W = patch.shape
if type == 0:#左上
return patch[H-valid_H:, W-valid_W:]
if type == 1:#右上
return patch[H-valid_H:, :valid_W]
if type == 2:#左下
return patch[:valid_H, W-valid_W:]
if type == 3:#右下
return patch[:valid_H, :valid_W]
return patch
def pose137_to_pose122(x):
return np.concatenate([x[:2, 0:8], # upper_body
x[:2, 15:17], # eyes
x[:2, 25:]], axis=1) # face, hand_l and hand_r
def scale_resize(curshape, myshape=(1080, 1920, 3), mean_height=0.0):
if curshape == myshape:
return None
x_mult = myshape[0] / float(curshape[0])
y_mult = myshape[1] / float(curshape[1])
if x_mult == y_mult:
# just need to scale
return x_mult, (0.0, 0.0)
elif y_mult > x_mult:
### scale x and center y
y_new = x_mult * float(curshape[1])
translate_y = (myshape[1] - y_new) / 2.0
return x_mult, (translate_y, 0.0)
### x_mult > y_mult
### already in landscape mode scale y, center x (rows)
x_new = y_mult * float(curshape[0])
translate_x = (myshape[0] - x_new) / 2.0
return y_mult, (0.0, translate_x)
def fix_scale_coords(points, scale, translate):
points = np.array(points).transpose(1,0)
points[0::3] = scale * points[0::3] + translate[0]
points[1::3] = scale * points[1::3] + translate[1]
return points.transpose(1,0)
if __name__=="__main__":
x,y = torch.meshgrid(torch.range(0,5),torch.range(0,5))
print(y)
a = torch.randn((5,3,64,64))
b = torch.randn((5,2,3,11))
# c = affine_warp(a, b)
mask = torch.randn((5,11,64,64))
a = make_warped_stack(mask,a,b)
print(a.shape)
repeat((torch.arange(2)*9).type(torch.FloatTensor), 5)