/
neural_style_field.py
309 lines (261 loc) · 15.5 KB
/
neural_style_field.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
import torch.nn as nn
import torch.nn.functional as F
import torch.optim
import torch
from utils import device
from sg_render import render_with_sg
from network import svbrdf_network
from network import Normal_estimation_network
import open3d as o3d
import numpy as np
import ipdb
width1 = 712
def get_rays(elev, azim, r=3.0,width = 512):
x = r * torch.cos(elev) * torch.cos(azim)
y = r * torch.sin(elev)
z = r * torch.cos(elev) * torch.sin(azim)
# print(elev,azim,x,y,z)
pos = np.array([x.numpy(),y.numpy(),z.numpy()])
look_at = np.array([-x.numpy(),-y.numpy(),-z.numpy()])
# direction = torch.tensor([0.0, 1.0, 0.0]).unsqueeze(0)
rays = o3d.t.geometry.RaycastingScene.create_rays_pinhole(fov_deg=60,
center=look_at,
eye=pos,
up=[0, -1, 0],
width_px=width,
height_px=width,
)
return rays
def get_rays1(elev, azim, r=3.0):
x = r * torch.cos(elev) * torch.cos(azim)
y = r * torch.sin(elev)
z = r * torch.cos(elev) * torch.sin(azim)
# print(elev,azim,x,y,z)
pos = np.array([x.numpy(),y.numpy(),z.numpy()])
look_at = np.array([-x.numpy(),-y.numpy(),-z.numpy()])
# direction = torch.tensor([0.0, 1.0, 0.0]).unsqueeze(0)
rays = o3d.t.geometry.RaycastingScene.create_rays_pinhole(fov_deg=60,
center=look_at,
eye=pos,
up=[0, -1, 0],
width_px=width1,
height_px=width1,
)
return rays
class NeuralStyleField(nn.Module):
def __init__(self,
material_random_pe_numfreq=0,
material_random_pe_sigma=12,
num_lgt_sgs=32,
max_delta_theta=np.pi/2,
max_delta_phi=np.pi/2,
normal_nerf_pe_numfreq=0,
normal_random_pe_numfreq=0,
symmetry=False,
radius=2.0,
background='black',
init_r_and_s=False,
width=512,
init_roughness=0.7,
init_specular=0.23,
material_nerf_pe_numfreq=0,
normal_random_pe_sigma=20,
if_normal_clamp=False):
"""_summary_
Args:
material_random_pe_numfreq (int, optional): the numer of frequencies using nerf's position encoding in svbrdf network. Defaults to 0.
material_random_pe_sigma (int, optional): the sigma of random position encoding in svbrdf network. Defaults to 12.
num_lgt_sgs (int, optional): the number of light SGs. Defaults to 32.
max_delta_theta (_type_, optional): maximum offset of elevation angle whose unit is radian. Defaults to np.pi/2.
max_delta_phi (_type_, optional): maximum offset of azimuth angle whose unit is radian. Defaults to np.pi/2.
normal_nerf_pe_numfreq (int, optional): the number of frequencies using nerf's position encoding in normal network. Defaults to 0.
normal_random_pe_numfreq (int, optional): the number of frequencies using random position encoding in normal network. Defaults to 0.
symmetry (bool, optional): With this symmetry prior, the texture of the mesh will be symmetrical along the z-axis.We use this parameter in person. Defaults to False.
radius (float, optional): the sampling raidus of camara position. Defaults to 2.0.
background (str, optional): the background of render image.'black','white' or 'gaussian' can be selected. Defaults to 'black'.
init_r_and_s (bool, optional): It will initialize roughness and specular if setting true. Defaults to False.
width (int, optional): the size of render image will be [width,width]. Defaults to 512.
init_roughness (float, optional): Initial value of roughness 0~1. Defaults to 0.7.
init_specular (float, optional): Initial value of specular 0~1. Defaults to 0.23.
material_nerf_pe_numfreq (int, optional): the numer of frequencies using nerf's position encoding in svbrdf network. Defaults to 0.
normal_random_pe_sigma (int, optional): the sigma of random position encoding in normal network. Defaults to 20.
if_normal_clamp (bool, optional): _description_. Defaults to False.
"""
super().__init__()
self.svbrdf_network = svbrdf_network(material_random_pe_numfreq = material_random_pe_numfreq,
material_random_pe_sigma = material_random_pe_sigma,
dim=256,
white_specular = False,
white_light = True,
num_lgt_sgs = num_lgt_sgs,
num_base_materials = 1,
upper_hemi = False,
fix_specular_albedo = False,
init_r_and_s = init_r_and_s,
init_roughness=init_roughness,
init_specular=init_specular,
material_nerf_pe_numfreq= material_nerf_pe_numfreq)
self.radius = radius
self.symmetry = symmetry
self.width = width
self.elev = 0.6283
self.azim = 0.5
self.Normal_estimation_network=Normal_estimation_network(max_delta_theta=max_delta_theta,
max_delta_phi=max_delta_phi,
normal_nerf_pe_numfreq=normal_nerf_pe_numfreq,
normal_random_pe_numfreq=normal_random_pe_numfreq,
normal_random_pe_sigma=normal_random_pe_sigma,
if_normal_clamp = if_normal_clamp)
if background=='black':
self.background = torch.zeros(width*width,3)
if background=='white':
self.background = torch.ones(width*width,3)
if background=='gaussian':
self.background = torch.randn(width*width,3)
self.background[:,1] =self.background[:,0]
self.background[:,2] =self.background[:,0]
def render_single_image(self, scene, azim , elev, r=2):
images = []
normal1 = []
normal2 = []
roughness = []
specular = []
diffuse = []
rays = get_rays1(elev, azim, r=r)
ans = scene.cast_rays(rays)
hit = ans['t_hit'].isfinite()
points = rays[hit][:,:3] + rays[hit][:,3:]*ans['t_hit'][hit].reshape((-1,1))
normal = ans['primitive_normals'][hit].reshape((-1,3))
view_dirs = -torch.nn.functional.normalize(torch.from_numpy(rays[hit][:,3:].numpy())).to(device)
pcd = o3d.t.geometry.PointCloud(points)
pcd.point["normals"] = normal
pcd = pcd.to_legacy()
points = torch.from_numpy(np.asarray(pcd.points)).float().to(device)
if self.symmetry:
points[:,2]=torch.abs(points[:,2])
normals1 = torch.nn.functional.normalize(torch.from_numpy(np.asarray(pcd.normals))).float().to(device)
normals2 = self.Normal_estimation_network(points,normals1)
normals2 = torch.nn.functional.normalize(normals2)
ret = self.get_rbg_value(points,normals2, view_dirs)
hit1 = torch.from_numpy(hit.reshape(width1*width1).numpy())
sg_rgb_values = torch.ones(width1*width1,3).float().to(device)
sg_rgb_values[hit1] = ret['sg_rgb']
sg_normal1_values = torch.ones(width1*width1,3).float().to(device)
sg_normal1_values[hit1]= normals1
sg_normal1_values = sg_normal1_values.reshape(1,width1,width1,3)
sg_normal2_values = torch.ones(width1*width1,3).float().to(device)
sg_normal2_values[hit1]= normals2
sg_normal2_values = sg_normal2_values.reshape(1,width1,width1,3)
sg_roughness_values = torch.ones(width1*width1,1).float().to(device)
sg_roughness_values[hit1] = ret['sg_roughness']
sg_roughness_values = sg_roughness_values.reshape(1,width1,width1,1)
sg_roughness_values = torch.clamp(sg_roughness_values, 0, 1)
sg_diffuse_values = torch.ones(width1*width1,3).float().to(device)
sg_diffuse_values[hit1] = ret['sg_diffuse_rgb']
sg_diffuse_values = sg_diffuse_values.reshape(1,width1,width1,3)
sg_diffuse_values = torch.clamp(sg_diffuse_values, 0, 1)
sg_specular_values = torch.ones(width1*width1,3).float().to(device)
sg_specular_values[hit1] = ret['sg_specular_rgb']
sg_specular_values = sg_specular_values.reshape(1,width1,width1,3)
sg_specular_values = torch.clamp(sg_specular_values, 0, 1)
normal1.append(sg_normal1_values)
normal2.append(sg_normal2_values)
roughness.append(sg_roughness_values)
diffuse.append(sg_diffuse_values)
specular.append(sg_specular_values)
image = sg_rgb_values.reshape(width1,width1,3).unsqueeze(0)
image = torch.clamp(image, 0, 1)
images.append(image)
images = torch.cat(images, dim=0).permute(0, 3, 1, 2)
normal1 = torch.cat(normal1, dim=0).permute(0, 3, 1, 2)
normal2 = torch.cat(normal2, dim=0).permute(0, 3, 1, 2)
roughness = torch.cat(roughness, dim=0).permute(0, 3, 1, 2)
diffuse = torch.cat(diffuse, dim=0).permute(0, 3, 1, 2)
specular = torch.cat(specular, dim=0).permute(0, 3, 1, 2)
return images, normal1, normal2, roughness, diffuse, specular
def forward(self, scene, num_views=8, std=8, center_elev=0, center_azim=0, eval=False, train_poses='default', train_render_elev=[5.6549, 0, 0.6283]):
# if num_views>1:
# self.elev = torch.cat((torch.tensor([center_elev]), torch.randn(num_views - 1) * np.pi / std + center_elev))
# self.azim = torch.cat((torch.tensor([center_azim]),torch.randn(num_views - 1) * 2 * np.pi / std + center_azim))
# if num_views==1:
# self.elev = torch.randn(num_views) * np.pi /std+ center_elev
# self.azim += torch.rand(num_views) * 0.1
if eval == False:
if train_poses == 'default':
if num_views>1:
self.elev = torch.cat((torch.tensor([center_elev]), torch.randn(num_views - 1) * np.pi / std + center_elev))
self.azim = torch.cat((torch.tensor([center_azim]),torch.randn(num_views - 1) * 2 * np.pi / std + center_azim))
if num_views==1:
self.elev = torch.randn(num_views) * np.pi /std+ center_elev
self.azim += torch.rand(num_views) * 0.1
elif train_poses == 'uniform': # three uniform
view_num = 100
self.azim = torch.linspace(0, 2 * np.pi + 0, view_num) # since 0 = 2π dont include last element
self.elev = torch.Tensor(train_render_elev) # -36 0 36
x = np.random.choice(range(view_num), size=num_views, replace=False)
y = np.random.choice(range(3), size=1)
self.azim = self.azim[x]
self.elev = self.elev[y]
else:
self.azim = [center_azim]
self.elev = [center_elev]
images_and_masks = []
depths = []
for i in range(num_views):
# rays = get_rays(self.elev[i], self.azim[i], r=self.radius,width=self.width)
rays = get_rays(self.elev[0], self.azim[i], r=self.radius,width=self.width)
ans = scene.cast_rays(rays)
hit = ans['t_hit'].isfinite()
points = rays[hit][:,:3] + rays[hit][:,3:]*ans['t_hit'][hit].reshape((-1,1))
normal = ans['primitive_normals'][hit].reshape((-1,3))
# get depth
dis = (points - rays[hit][:,:3]).numpy()
norm_dis = np.linalg.norm(dis, axis=1)
# background 2*max(dis)
depth = np.zeros((rays.shape[0], rays.shape[1]))
# depth[:,:] = 2 * np.max(norm_dis)
# depth[:,:] = np.max(norm_dis) + 0.1 * np.max(norm_dis)
hit_index = np.nonzero(hit.numpy())
depth[hit_index[0], hit_index[1]] = norm_dis
# depth normalizaton
depth -= np.min(depth)
depth /= np.max(depth)
depth = (depth * 255.0).clip(0, 255)
depth = depth[:, :, None] # (512, 512, 1)
save_depth = np.concatenate([depth, depth, depth], axis=2)
depths.append(save_depth)
# import ipdb
# ipdb.set_trace()
view_dirs = -torch.nn.functional.normalize(torch.from_numpy(rays[hit][:,3:].numpy())).to(device)
pcd = o3d.t.geometry.PointCloud(points)
pcd.point["normals"] = normal
pcd = pcd.to_legacy()
points = torch.from_numpy(np.asarray(pcd.points)).float().to(device)
if self.symmetry:
points[:,2]=torch.abs(points[:,2])
normals1= torch.nn.functional.normalize(torch.from_numpy(np.asarray(pcd.normals))).float().to(device)
normals2 = self.Normal_estimation_network(points,normals1)
normals2 = torch.nn.functional.normalize(normals2)
ret = self.get_rbg_value(points, normals2, view_dirs)
hit1 = torch.from_numpy(hit.reshape(self.width*self.width).numpy())
sg_rgb_values = self.background.float().to(device)
sg_rgb_values[hit1] = ret['sg_rgb']
mask = torch.from_numpy(hit.numpy()).float().cuda().reshape(1,self.width,self.width,1)
image = sg_rgb_values.reshape(self.width,self.width,3).unsqueeze(0)
image = torch.clamp(image, 0, 1)
image_and_mask = torch.cat((image,mask),dim=3)
images_and_masks.append(image_and_mask)
images_and_masks = torch.cat(images_and_masks, dim=0).permute(0, 3, 1, 2)
depths = np.array(depths, dtype=np.float32)
return images_and_masks, depths, self.azim, self.elev
def get_rbg_value(self, points,normals, view_dirs):
ret = { }
sg_envmap_material = self.svbrdf_network(points)
sg_ret = render_with_sg(lgtSGs=sg_envmap_material['sg_lgtSGs'],
specular_reflectance=sg_envmap_material['sg_specular_reflectance'],
roughness=sg_envmap_material['sg_roughness'],
diffuse_albedo=sg_envmap_material['sg_diffuse_albedo'],
normal=normals, viewdirs=view_dirs,
)
ret.update(sg_ret)
return ret