-
Notifications
You must be signed in to change notification settings - Fork 6
/
utils.py
226 lines (191 loc) · 7.88 KB
/
utils.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
import torch
import numpy as np
import pickle as pkl
from pathlib import Path
import subprocess
import random
from nvdiff_render import util
import torchvision
import torchvision.transforms as T
downsampler_512 = T.Resize((512, 512))
tensor_to_img = T.ToPILImage()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
SMPL_UV_OBJ_PATH = './data/smpl_uv.obj'
def sample_view_obj(n_view, cam_radius, res=[512, 512], cam_near_far=[0.1, 1000.0], spp=1, is_face=False):
iter_res = res
fovy = np.deg2rad(45)
proj_mtx = util.perspective(fovy, iter_res[1] / iter_res[0], cam_near_far[0], cam_near_far[1])
# Random rotation/translation matrix for optimization.
mv_list, mvp_list, campos_list, direction_list = [], [], [], []
for view_i in range(n_view):
if view_i == 0:
angle_x = 0.0 # elevation
angle_y = 0.0 # azimuth
else:
angle_x = np.random.uniform(-np.pi / 3, np.pi / 3)
angle_y = np.random.uniform(0, 2 * np.pi)
# direction
# 0 = front, 1 = side, 2 = back, 3 = overhead
if angle_x < -np.pi / 4:
direction = 3
else:
if 0 <= angle_y <= np.pi / 4 or angle_y > 7 * np.pi / 4:
direction = 0
elif np.pi / 4 < angle_y <= 3 * np.pi / 4:
direction = 1
elif 3 * np.pi / 4 < angle_y <= 5 * np.pi / 4:
direction = 2
elif 5 * np.pi / 4 < angle_y <= 7 * np.pi / 4:
direction = 1
# for object, hard to tell front, back. so, perform prompt augment for only overhead view
# If the results do not look good, you may use this direction prompts.
# if angle_x < -np.pi / 4:
# direction = 1
# else:
# direction = 0
mv = util.translate(0, 0, -cam_radius) @ (util.rotate_x(angle_x) @ util.rotate_y(angle_y))
mvp = proj_mtx @ mv
campos = torch.linalg.inv(mv)[:3, 3]
mv_list.append(mv[None, ...].cuda())
mvp_list.append(mvp[None, ...].cuda())
campos_list.append(campos[None, ...].cuda())
direction_list.append(direction)
cam = {
'mv': torch.cat(mv_list, dim=0),
'mvp': torch.cat(mvp_list, dim=0),
'campos': torch.cat(campos_list, dim=0),
'direction': np.array(direction_list, dtype=np.int32),
'resolution': iter_res,
'spp': spp
}
return cam
def sample_view_human(n_view, cam_radius, res=[512, 512], cam_near_far=[0.1, 1000.0], spp=1, is_face=False):
iter_res = res
fovy = np.deg2rad(45)
proj_mtx = util.perspective(fovy, iter_res[1] / iter_res[0], cam_near_far[0], cam_near_far[1])
# ==============================================================================================
# Random camera & light position
# ==============================================================================================
# Random rotation/translation matrix for optimization.
mv_list, mvp_list, campos_list, direction_list = [], [], [], []
for view_i in range(n_view):
if is_face:
angle_x = np.random.uniform(-np.pi / 3, np.pi / 6)
angle_y = np.random.uniform(-np.pi / 2, np.pi / 2)
else:
angle_x = np.random.uniform(-np.pi / 3, np.pi / 3)
angle_y = np.random.uniform(0, 2 * np.pi)
# direction
# 0 = front, 1 = side, 2 = back, 3 = overhead
if not is_face:
if angle_x < -np.pi / 4:
direction = 3
else:
if angle_y >= 0 and angle_y <= np.pi / 8 or angle_y > 15 * np.pi / 8:
direction = 0
elif angle_y > np.pi / 8 and angle_y <= 7 * np.pi / 8:
direction = 1
elif angle_y > 7 * np.pi / 8 and angle_y <= 9 * np.pi / 8:
direction = 2
elif angle_y > 9 * np.pi / 8 and angle_y <= 15 * np.pi / 8:
direction = 1
else:
if angle_x < -np.pi / 4:
direction = 2
else:
if -np.pi / 8 <= angle_y < np.pi / 8:
direction = 0
else:
direction = 1
mv = util.translate(0, 0, -cam_radius) @ (util.rotate_x(angle_x) @ util.rotate_y(angle_y))
mvp = proj_mtx @ mv
campos = torch.linalg.inv(mv)[:3, 3]
mv_list.append(mv[None, ...].cuda())
mvp_list.append(mvp[None, ...].cuda())
campos_list.append(campos[None, ...].cuda())
direction_list.append(direction)
cam = {
'mv': torch.cat(mv_list, dim=0),
'mvp': torch.cat(mvp_list, dim=0),
'campos': torch.cat(campos_list, dim=0),
'direction': np.array(direction_list, dtype=np.int32),
'resolution': iter_res,
'spp': spp
}
return cam
def sample_circle_view(n_view, elev, cam_radius, res=[512, 512], cam_near_far=[0.1, 1000.0], spp=1):
iter_res = res
fovy = np.deg2rad(45)
proj_mtx = util.perspective(fovy, iter_res[1] / iter_res[0], cam_near_far[0], cam_near_far[1])
# Random rotation/translation matrix for optimization.
mv_list, mvp_list, campos_list, direction_list = [], [], [], []
angles_y = np.linspace(0.0, 2 * np.pi, n_view)
for view_i in range(n_view):
angle_x = elev
angle_y = angles_y[view_i]
# direction
# 0 = front, 1 = side, 2 = back, 3 = side
if angle_y >= 0 and angle_y <= np.pi / 8 or angle_y > 15 * np.pi / 8:
direction = 0
elif angle_y > np.pi / 8 and angle_y <= 7 * np.pi / 8:
direction = 1
elif angle_y > 7 * np.pi / 8 and angle_y <= 9 * np.pi / 8:
direction = 2
elif angle_y > 9 * np.pi / 8 and angle_y <= 15 * np.pi / 8:
direction = 3
mv = util.translate(0, 0, -cam_radius) @ (util.rotate_x(angle_x) @ util.rotate_y(angle_y))
mvp = proj_mtx @ mv
campos = torch.linalg.inv(mv)[:3, 3]
mv_list.append(mv[None, ...].cuda())
mvp_list.append(mvp[None, ...].cuda())
campos_list.append(campos[None, ...].cuda())
direction_list.append(direction)
cam = {
'mv': torch.cat(mv_list, dim=0),
'mvp': torch.cat(mvp_list, dim=0),
'campos': torch.cat(campos_list, dim=0),
'direction': np.array(direction_list, dtype=np.int32),
'resolution': iter_res,
'spp': spp
}
return cam
def create_video(img_path, out_path, fps=60):
'''
Creates a video from the frame format in the given directory and saves to out_path.
'''
command = ['/usr/bin/ffmpeg', '-y', '-r', str(fps), '-i', img_path, \
'-vcodec', 'libx264', '-crf', '25', '-pix_fmt', 'yuv420p', out_path]
subprocess.run(command)
def imgcat(x, renormalize=False):
# x: [3, H, W] or [1, H, W] or [H, W]
import matplotlib.pyplot as plt
import numpy as np
import torch
if isinstance(x, torch.Tensor):
if len(x.shape) == 3:
x = x.permute(1, 2, 0).squeeze()
x = x.detach().cpu().numpy()
print(f'[imgcat] {x.shape}, {x.dtype}, {x.min()} ~ {x.max()}')
x = x.astype(np.float32)
# renormalize
if renormalize:
x = (x - x.min(axis=0, keepdims=True)) / (x.max(axis=0, keepdims=True) - x.min(axis=0, keepdims=True) + 1e-8)
plt.imshow(x)
plt.show()
from pytorch3d.transforms import axis_angle_to_matrix, matrix_to_axis_angle
from pytorch3d.structures import Meshes
from pytorch3d.io import load_obj, save_obj
#
def load_obj_uv(obj_path, device):
vert, face, aux = load_obj(obj_path, device=device)
vt = aux.verts_uvs
ft = face.textures_idx
vt = torch.cat((vt[:, [0]], 1.0 - vt[:, [1]]), dim=1)
return ft, vt, face.verts_idx, vert
def load_smpl_uv(device):
# smpl
vert, face, aux = load_obj(SMPL_UV_OBJ_PATH, device=device)
vt = aux.verts_uvs
ft = face.textures_idx
vt = torch.cat((vt[:, [0]], 1.0 - vt[:, [1]]), dim=1)
return ft, vt