-
Notifications
You must be signed in to change notification settings - Fork 2.3k
/
mitsuba_material_estimation.py
236 lines (203 loc) · 8.63 KB
/
mitsuba_material_estimation.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
# ----------------------------------------------------------------------------
# - Open3D: www.open3d.org -
# ----------------------------------------------------------------------------
# Copyright (c) 2018-2023 www.open3d.org
# SPDX-License-Identifier: MIT
# ----------------------------------------------------------------------------
import sys
import argparse
from pathlib import Path
import open3d as o3d
import mitsuba as mi
import drjit as dr
import numpy as np
import math
def make_mitsuba_scene(mesh, cam_xform, fov, width, height, principle_pts,
envmap):
# Camera transform
t_from_np = mi.ScalarTransform4f(cam_xform)
# Transform necessary to get from Open3D's environment map coordinate system
# to Mitsuba's
env_t = mi.ScalarTransform4f.rotate(axis=[0, 0, 1],
angle=90).rotate(axis=[1, 0, 0],
angle=90)
scene_dict = {
"type": "scene",
"integrator": {
'type': 'path'
},
"light": {
"type": "envmap",
"to_world": env_t,
"bitmap": mi.Bitmap(envmap),
},
"sensor": {
"type": "perspective",
"fov": fov,
"to_world": t_from_np,
"principal_point_offset_x": principle_pts[0],
"principal_point_offset_y": principle_pts[1],
"thefilm": {
"type": "hdrfilm",
"width": width,
"height": height,
},
"thesampler": {
"type": "multijitter",
"sample_count": 64,
},
},
"themesh": mesh,
}
scene = mi.load_dict(scene_dict)
return scene
def run_estimation(mesh, cam_info, ref_image, env_width, iterations, tv_alpha):
# Make Mitsuba mesh from Open3D mesh -- conversion will attach a Mitsuba
# Principled BSDF to the mesh
mesh_opt = mesh.to_mitsuba('themesh')
# Prepare empty environment map
empty_envmap = np.ones((int(env_width / 2), env_width, 3))
# Create Mitsuba scene
scene = make_mitsuba_scene(mesh_opt, cam_info[0], cam_info[1], cam_info[2],
cam_info[3], cam_info[4], empty_envmap)
def total_variation(image, alpha):
diff1 = image[1:, :, :] - image[:-1, :, :]
diff2 = image[:, 1:, :] - image[:, :-1, :]
return alpha * (dr.sum(dr.abs(diff1)) / len(diff1) +
dr.sum(dr.abs(diff2)) / len(diff2))
def mse(image, ref_img):
return dr.mean(dr.sqr(image - ref_img))
params = mi.traverse(scene)
print(params)
# Create a Mitsuba Optimizer and configure it to optimize albedo and
# environment maps
opt = mi.ad.Adam(lr=0.05, mask_updates=True)
opt['themesh.bsdf.base_color.data'] = params['themesh.bsdf.base_color.data']
opt['light.data'] = params['light.data']
params.update(opt)
integrator = mi.load_dict({'type': 'prb'})
for i in range(iterations):
img = mi.render(scene, params, spp=8, seed=i, integrator=integrator)
# Compute loss
loss = mse(img, ref_image)
# Apply TV regularization if requested
if tv_alpha > 0.0:
loss = loss + total_variation(opt['themesh.bsdf.base_color.data'],
tv_alpha)
# Backpropogate and step. Note: if we were optimizing over a larger set
# of inputs not just a single image we might want to step only every x
# number of inputs
dr.backward(loss)
opt.step()
# Make sure albedo values stay in allowed range
opt['themesh.bsdf.base_color.data'] = dr.clamp(
opt['themesh.bsdf.base_color.data'], 0.0, 1.0)
params.update(opt)
print(f'Iteration {i} complete')
# Done! Return the estimated maps
albedo_img = params['themesh.bsdf.base_color.data'].numpy()
envmap_img = params['light.data'].numpy()
return (albedo_img, envmap_img)
def load_input_mesh(model_path, tex_dim):
mesh = o3d.t.io.read_triangle_mesh(str(model_path))
mesh.material.set_default_properties()
mesh.material.material_name = 'defaultLit' # note: ignored by Mitsuba, just used to visualize in Open3D
mesh.material.texture_maps['albedo'] = o3d.t.geometry.Image(0.5 + np.zeros(
(tex_dim, tex_dim, 3), dtype=np.float32))
return mesh
def load_input_data(object, camera_pose, input_image, tex_dim):
print(f'Loading {object}...')
mesh = load_input_mesh(object, tex_dim)
print(f'Loading camera pose from {camera_pose}...')
cam_npz = np.load(camera_pose)
img_width = cam_npz['width'].item()
img_height = cam_npz['height'].item()
cam_xform = np.linalg.inv(cam_npz['T'])
cam_xform = np.matmul(
cam_xform,
np.array([[-1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]],
dtype=np.float32))
fov = 2 * np.arctan(0.5 * img_width / cam_npz['K'][0, 0])
fov = (180.0 / math.pi) * fov.item()
camera = (cam_xform, fov, img_width, img_height, (0.0, 0.0))
print(f'Loading reference image from {input_image}...')
ref_img = o3d.t.io.read_image(str(input_image))
ref_img = ref_img.as_tensor()[:, :, 0:3].to(o3d.core.Dtype.Float32) / 255.0
bmp = mi.Bitmap(ref_img.numpy()).convert(srgb_gamma=False)
ref_img = mi.TensorXf(bmp)
return (mesh, camera, ref_img)
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description=
"Script that estimates texture and environment map from an input image and geometry. You can find data to test this script here: https://github.com/isl-org/open3d_downloads/releases/download/mitsuba-demos/raven_mitsuba.zip.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
'object_path',
type=Path,
help=
"Path to geometry for which to estimate albedo. It is assumed that in the same directory will be an object-name.npz which contains the camera pose information and an object-name.png which is the input image"
)
parser.add_argument('--env-width', type=int, default=1024)
parser.add_argument('--tex-width',
type=int,
default=2048,
help="The dimensions of the texture")
parser.add_argument(
'--device',
default='cuda' if o3d.core.cuda.is_available() else 'cpu',
choices=('cpu', 'cuda'),
help="Run Mitsuba on 'cuda' or 'cpu'")
parser.add_argument('--iterations',
type=int,
default=40,
help="Number of iterations")
parser.add_argument(
'--total-variation',
type=float,
default=0.01,
help="Factor to apply to total_variation loss. 0.0 disables TV")
if len(sys.argv) < 2:
parser.print_help(sys.stderr)
sys.exit(1)
args = parser.parse_args()
print("Arguments: ", vars(args))
# Initialize Mitsuba
if args.device == 'cpu':
mi.set_variant('llvm_ad_rgb')
else:
mi.set_variant('cuda_ad_rgb')
# Confirm that the 3 required inputs exist
object_path = args.object_path
object_name = object_path.stem
datadir = args.object_path.parent
camera_pose = datadir / (object_name + '.npz')
input_image = datadir / (object_name + '.png')
if not object_path.exists():
print(f'{object_path} does not exist!')
sys.exit()
if not camera_pose.exists():
print(f'{camera_pose} does not exist!')
sys.exit()
if not input_image.exists():
print(f'{input_image} does not exist!')
sys.exit()
# Load input data
mesh, cam_info, input_image = load_input_data(object_path, camera_pose,
input_image, args.tex_width)
# Estimate albedo map
print('Running material estimation...')
albedo, envmap = run_estimation(mesh, cam_info, input_image, args.env_width,
args.iterations, args.total_variation)
# Save maps
def save_image(img, name, output_dir):
# scale to 0-255
texture = o3d.core.Tensor(img * 255.0).to(o3d.core.Dtype.UInt8)
texture = o3d.t.geometry.Image(texture)
o3d.t.io.write_image(str(output_dir / name), texture)
print('Saving final results...')
save_image(albedo, 'estimated_albedo.png', datadir)
mi.Bitmap(envmap).write(str(datadir / 'predicted_envmap.exr'))
# Visualize result with Open3D
mesh.material.texture_maps['albedo'] = o3d.t.io.read_image(
str(datadir / 'estimated_albedo.png'))
o3d.visualization.draw(mesh)