-
Notifications
You must be signed in to change notification settings - Fork 4
/
ReasonOcclusionPyTorch3D.py
154 lines (103 loc) · 6.21 KB
/
ReasonOcclusionPyTorch3D.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
import torch
import numpy as np
from pytorch3d.renderer import PerspectiveCameras, look_at_view_transform, MeshRasterizer, RasterizationSettings
from pytorch3d.structures import Meshes
from pytorch3d.renderer.mesh.utils import interpolate_face_attributes
import matplotlib.pyplot as plt
import imageio
from pytorch3d.renderer.blending import softmax_rgb_blend, hard_rgb_blend, BlendParams
from PIL import Image
from VoGE.Converter import Cuboid
import os
def interpolate_attributes_as_img(fragments, face_colors, **kwargs):
colors_img = interpolate_face_attributes(fragments.pix_to_face, fragments.bary_coords, face_colors)
blend_params = BlendParams(**kwargs)
if colors_img.shape[3] == 1:
return hard_rgb_blend(colors_img, fragments, blend_params).squeeze()
else:
return softmax_rgb_blend(colors_img, fragments, blend_params).squeeze()
def save_image(get, save_name):
get = get.clamp(min=0, max=1)[..., :3].detach().cpu().numpy() * 255
Image.fromarray(get.astype(np.uint8)).save(save_name)
if __name__ == '__main__':
device = 'cuda'
colors_ = np.array([[0, 0.2, 1], [0, 0.2, 1], [0, 1, 0.2], [0, 1, 0.2], [0, 1, 1], [0, 1, 1]])
verts, faces, colors = Cuboid.cuboid_mesh((-0.8, 0.8), (-0.4, 0.4),(-0.6, 0.6), 4000, colors=colors_)
verts = torch.from_numpy(verts).type(torch.float32).to(device)
faces = torch.from_numpy(faces).type(torch.long).to(device)
colors = torch.from_numpy(colors).type(torch.float32).to(device)
faces_color0 = colors[faces]
colors_ = np.array([[1, 0.2, 0], [1, 0.2, 0], [1, 1, 0], [1, 1, 0], [0.2, 1, 0], [0.2, 1, 0]])
verts1, faces1, colors1 = Cuboid.cuboid_mesh((-1, 1), (-1, 1), (-0.3, 0.3), 3000, colors=colors_)
n_v0, n_v1 = verts.shape[0], verts1.shape[0]
verts1 = torch.from_numpy(verts1).type(torch.float32).to(device)
faces1 = torch.from_numpy(faces1).type(torch.long).to(device)
colors1 = torch.from_numpy(colors1).type(torch.float32).to(device)
v_base0 = torch.Tensor([[0.5, 0, 1]]).to(device)
v_base1 = torch.Tensor([[0, 0, 0]]).to(device)
faces_color1 = colors1[faces1]
faces_colors = torch.cat((faces_color0, faces_color1), dim=0)
R, T = look_at_view_transform(dist=5, elev=10, azim=20, device=device)
R1, T1 = look_at_view_transform(dist=6, elev=10, azim=-50, device=device)
T1 += torch.Tensor([[-1, 0, 0]]).to(device)
image_size = (400, 400)
save_head = 'reason_occ/SoftRas'
save_pendix = '_1'
os.makedirs('reason_occ', exist_ok=True)
# The default gamma is 1e-4, however, which make it hardly update.
# we have tried 1e-2, which makes it converge to some other local min
faces_per_pixel = 60
sigmas = 1e-3
gamma = 1e-3
camera = PerspectiveCameras(focal_length=300, in_ndc=False, device=device, image_size=(image_size, ), principal_point=((image_size[0] // 2, image_size[1] // 2), ), T=T, R=R)
raster_settings = RasterizationSettings(image_size=image_size, faces_per_pixel=faces_per_pixel, blur_radius=sigmas)
raster = MeshRasterizer(cameras=camera, raster_settings=raster_settings)
meshes = Meshes(verts=[verts, verts1], faces=[faces, faces1])
meshes = Meshes(verts=[meshes.verts_packed()], faces=[meshes.faces_packed()])
off_set = torch.cat((v_base0.expand(n_v0, -1), v_base1.expand(n_v1, -1)), dim=0)
new_meshes = meshes.offset_verts(off_set)
frag = raster(new_meshes)
timg = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(0., 0., 0.)).detach()
v_pred0 = torch.Tensor([[-1, 0, -5]]).to(device)
v_pred1 = torch.Tensor([[0, 0, 0]]).to(device)
v_pred1.requires_grad = True
v_pred0.requires_grad = True
optimizer = torch.optim.Adam([v_pred0, v_pred1], lr=0.05, betas=(0.6, 0.4))
l2loss = torch.nn.MSELoss()
with torch.no_grad():
off_set = torch.cat((v_pred0.expand(n_v0, -1), v_pred1.expand(n_v1, -1)), dim=0)
new_meshes = meshes.offset_verts(off_set)
frag = raster(new_meshes, R=R, T=T)
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(1., 1., 1.))
save_image(img, save_head + '_before_V0' + save_pendix + '.png')
save_image(timg, save_head + '_target' + save_pendix + '.png')
frag = raster(new_meshes, R=R1, T=T1)
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(1., 1., 1.))
save_image(img, save_head + '_before_V1' + save_pendix + '.png')
out_image = []
for iter_ in range(200):
off_set = torch.cat((v_pred0.expand(n_v0, -1), v_pred1.expand(n_v1, -1)), dim=0)
new_meshes = meshes.offset_verts(off_set)
frag = raster(new_meshes, R=R, T=T)
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(0., 0., 0.))
loss_ = l2loss(img, timg)
loss_.backward()
optimizer.step()
optimizer.zero_grad()
print(v_pred0)
im_np = torch.cat((img, timg), dim=1).clamp(min=0, max=1).detach().cpu().numpy() * 255
out_image.append(im_np.astype(np.uint8))
imageio.mimwrite(save_head + '_2cuboid%s.mp4' % save_pendix, out_image, fps=24, quality=8)
with torch.no_grad():
off_set = torch.cat((v_pred0.expand(n_v0, -1), v_pred1.expand(n_v1, -1)), dim=0)
new_meshes = meshes.offset_verts(off_set)
frag = raster(new_meshes, R=R, T=T)
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(1., 1., 1.))
save_image(img, save_head + '_after_V0' + save_pendix + '.png')
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(0., 0., 0.))
save_image(img, save_head + '_optimized' + save_pendix + '.png')
frag = raster(new_meshes, R=R1, T=T1)
img = interpolate_attributes_as_img(frag, faces_colors, sigma=sigmas, gamma=gamma, background_color=(1., 1., 1.))
save_image(img, save_head + '_after_V1' + save_pendix + '.png')
# plt.imshow(timg.detach().cpu().numpy())
# plt.show()