move part of the demo to the lib
Browse files- demo.py +2 -267
- dust3r +1 -1
- mast3r/demo.py +277 -0
demo.py
CHANGED
@@ -3,286 +3,21 @@
|
|
3 |
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
|
4 |
#
|
5 |
# --------------------------------------------------------
|
6 |
-
# gradio demo
|
7 |
# --------------------------------------------------------
|
8 |
-
import math
|
9 |
-
import gradio
|
10 |
import os
|
11 |
import torch
|
12 |
-
import numpy as np
|
13 |
import tempfile
|
14 |
-
import functools
|
15 |
-
import trimesh
|
16 |
-
import copy
|
17 |
-
from scipy.spatial.transform import Rotation
|
18 |
|
19 |
-
from mast3r.
|
20 |
-
from mast3r.cloud_opt.tsdf_optimizer import TSDFPostProcess
|
21 |
|
22 |
from mast3r.model import AsymmetricMASt3R
|
23 |
from mast3r.utils.misc import hash_md5
|
24 |
-
import mast3r.utils.path_to_dust3r # noqa
|
25 |
-
from dust3r.image_pairs import make_pairs
|
26 |
-
from dust3r.utils.image import load_images
|
27 |
-
from dust3r.utils.device import to_numpy
|
28 |
-
from dust3r.viz import add_scene_cam, CAM_COLORS, OPENGL, pts3d_to_trimesh, cat_meshes
|
29 |
-
from dust3r.demo import get_args_parser as dust3r_get_args_parser
|
30 |
|
31 |
import matplotlib.pyplot as pl
|
32 |
pl.ion()
|
33 |
|
34 |
torch.backends.cuda.matmul.allow_tf32 = True # for gpu >= Ampere and pytorch >= 1.12
|
35 |
-
batch_size = 1
|
36 |
-
|
37 |
-
|
38 |
-
def get_args_parser():
|
39 |
-
parser = dust3r_get_args_parser()
|
40 |
-
parser.add_argument('--share', action='store_true')
|
41 |
-
|
42 |
-
actions = parser._actions
|
43 |
-
for action in actions:
|
44 |
-
if action.dest == 'model_name':
|
45 |
-
action.choices = ["MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"]
|
46 |
-
# change defaults
|
47 |
-
parser.prog = 'mast3r demo'
|
48 |
-
return parser
|
49 |
-
|
50 |
-
|
51 |
-
def _convert_scene_output_to_glb(outdir, imgs, pts3d, mask, focals, cams2world, cam_size=0.05,
|
52 |
-
cam_color=None, as_pointcloud=False,
|
53 |
-
transparent_cams=False, silent=False):
|
54 |
-
assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)
|
55 |
-
pts3d = to_numpy(pts3d)
|
56 |
-
imgs = to_numpy(imgs)
|
57 |
-
focals = to_numpy(focals)
|
58 |
-
cams2world = to_numpy(cams2world)
|
59 |
-
|
60 |
-
scene = trimesh.Scene()
|
61 |
-
|
62 |
-
# full pointcloud
|
63 |
-
if as_pointcloud:
|
64 |
-
pts = np.concatenate([p[m.ravel()] for p, m in zip(pts3d, mask)])
|
65 |
-
col = np.concatenate([p[m] for p, m in zip(imgs, mask)])
|
66 |
-
pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))
|
67 |
-
scene.add_geometry(pct)
|
68 |
-
else:
|
69 |
-
meshes = []
|
70 |
-
for i in range(len(imgs)):
|
71 |
-
meshes.append(pts3d_to_trimesh(imgs[i], pts3d[i].reshape(imgs[i].shape), mask[i]))
|
72 |
-
mesh = trimesh.Trimesh(**cat_meshes(meshes))
|
73 |
-
scene.add_geometry(mesh)
|
74 |
-
|
75 |
-
# add each camera
|
76 |
-
for i, pose_c2w in enumerate(cams2world):
|
77 |
-
if isinstance(cam_color, list):
|
78 |
-
camera_edge_color = cam_color[i]
|
79 |
-
else:
|
80 |
-
camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]
|
81 |
-
add_scene_cam(scene, pose_c2w, camera_edge_color,
|
82 |
-
None if transparent_cams else imgs[i], focals[i],
|
83 |
-
imsize=imgs[i].shape[1::-1], screen_width=cam_size)
|
84 |
-
|
85 |
-
rot = np.eye(4)
|
86 |
-
rot[:3, :3] = Rotation.from_euler('y', np.deg2rad(180)).as_matrix()
|
87 |
-
scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))
|
88 |
-
outfile = os.path.join(outdir, 'scene.glb')
|
89 |
-
if not silent:
|
90 |
-
print('(exporting 3D scene to', outfile, ')')
|
91 |
-
scene.export(file_obj=outfile)
|
92 |
-
return outfile
|
93 |
-
|
94 |
-
|
95 |
-
def get_3D_model_from_scene(outdir, silent, scene, min_conf_thr=2, as_pointcloud=False, mask_sky=False,
|
96 |
-
clean_depth=False, transparent_cams=False, cam_size=0.05, TSDF_thresh=0):
|
97 |
-
"""
|
98 |
-
extract 3D_model (glb file) from a reconstructed scene
|
99 |
-
"""
|
100 |
-
if scene is None:
|
101 |
-
return None
|
102 |
-
|
103 |
-
# get optimized values from scene
|
104 |
-
rgbimg = scene.imgs
|
105 |
-
focals = scene.get_focals().cpu()
|
106 |
-
cams2world = scene.get_im_poses().cpu()
|
107 |
-
|
108 |
-
# 3D pointcloud from depthmap, poses and intrinsics
|
109 |
-
if TSDF_thresh > 0:
|
110 |
-
tsdf = TSDFPostProcess(scene, TSDF_thresh=TSDF_thresh)
|
111 |
-
pts3d, _, confs = to_numpy(tsdf.get_dense_pts3d(clean_depth=clean_depth))
|
112 |
-
else:
|
113 |
-
pts3d, _, confs = to_numpy(scene.get_dense_pts3d(clean_depth=clean_depth))
|
114 |
-
msk = to_numpy([c > min_conf_thr for c in confs])
|
115 |
-
return _convert_scene_output_to_glb(outdir, rgbimg, pts3d, msk, focals, cams2world, as_pointcloud=as_pointcloud,
|
116 |
-
transparent_cams=transparent_cams, cam_size=cam_size, silent=silent)
|
117 |
-
|
118 |
-
|
119 |
-
def get_reconstructed_scene(outdir, model, device, silent, image_size, filelist, optim_level, lr1, niter1, lr2, niter2,
|
120 |
-
min_conf_thr, matching_conf_thr, as_pointcloud, mask_sky, clean_depth, transparent_cams,
|
121 |
-
cam_size, scenegraph_type, winsize, win_cyclic, refid, TSDF_thresh, shared_intrinsics,
|
122 |
-
**kw):
|
123 |
-
"""
|
124 |
-
from a list of images, run mast3r inference, sparse global aligner.
|
125 |
-
then run get_3D_model_from_scene
|
126 |
-
"""
|
127 |
-
imgs = load_images(filelist, size=image_size, verbose=not silent)
|
128 |
-
if len(imgs) == 1:
|
129 |
-
imgs = [imgs[0], copy.deepcopy(imgs[0])]
|
130 |
-
imgs[1]['idx'] = 1
|
131 |
-
filelist = [filelist[0], filelist[0] + '_2']
|
132 |
-
|
133 |
-
scene_graph_params = [scenegraph_type]
|
134 |
-
if scenegraph_type in ["swin", "logwin"]:
|
135 |
-
scene_graph_params.append(str(winsize))
|
136 |
-
elif scenegraph_type == "oneref":
|
137 |
-
scene_graph_params.append(str(refid))
|
138 |
-
if scenegraph_type in ["swin", "logwin"] and not win_cyclic:
|
139 |
-
scene_graph_params.append('noncyclic')
|
140 |
-
scene_graph = '-'.join(scene_graph_params)
|
141 |
-
pairs = make_pairs(imgs, scene_graph=scene_graph, prefilter=None, symmetrize=True)
|
142 |
-
if optim_level == 'coarse':
|
143 |
-
niter2 = 0
|
144 |
-
# Sparse GA (forward mast3r -> matching -> 3D optim -> 2D refinement -> triangulation)
|
145 |
-
scene = sparse_global_alignment(filelist, pairs, os.path.join(outdir, 'cache'),
|
146 |
-
model, lr1=lr1, niter1=niter1, lr2=lr2, niter2=niter2, device=device,
|
147 |
-
opt_depth='depth' in optim_level, shared_intrinsics=shared_intrinsics,
|
148 |
-
matching_conf_thr=matching_conf_thr, **kw)
|
149 |
-
outfile = get_3D_model_from_scene(outdir, silent, scene, min_conf_thr, as_pointcloud, mask_sky,
|
150 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh)
|
151 |
-
return scene, outfile
|
152 |
-
|
153 |
-
|
154 |
-
def set_scenegraph_options(inputfiles, win_cyclic, refid, scenegraph_type):
|
155 |
-
num_files = len(inputfiles) if inputfiles is not None else 1
|
156 |
-
show_win_controls = scenegraph_type in ["swin", "logwin"]
|
157 |
-
show_winsize = scenegraph_type in ["swin", "logwin"]
|
158 |
-
show_cyclic = scenegraph_type in ["swin", "logwin"]
|
159 |
-
max_winsize, min_winsize = 1, 1
|
160 |
-
if scenegraph_type == "swin":
|
161 |
-
if win_cyclic:
|
162 |
-
max_winsize = max(1, math.ceil((num_files - 1) / 2))
|
163 |
-
else:
|
164 |
-
max_winsize = num_files - 1
|
165 |
-
elif scenegraph_type == "logwin":
|
166 |
-
if win_cyclic:
|
167 |
-
half_size = math.ceil((num_files - 1) / 2)
|
168 |
-
max_winsize = max(1, math.ceil(math.log(half_size, 2)))
|
169 |
-
else:
|
170 |
-
max_winsize = max(1, math.ceil(math.log(num_files, 2)))
|
171 |
-
winsize = gradio.Slider(label="Scene Graph: Window Size", value=max_winsize,
|
172 |
-
minimum=min_winsize, maximum=max_winsize, step=1, visible=show_winsize)
|
173 |
-
win_cyclic = gradio.Checkbox(value=win_cyclic, label="Cyclic sequence", visible=show_cyclic)
|
174 |
-
win_col = gradio.Column(visible=show_win_controls)
|
175 |
-
refid = gradio.Slider(label="Scene Graph: Id", value=0, minimum=0,
|
176 |
-
maximum=num_files - 1, step=1, visible=scenegraph_type == 'oneref')
|
177 |
-
return win_col, winsize, win_cyclic, refid
|
178 |
-
|
179 |
-
|
180 |
-
def main_demo(tmpdirname, model, device, image_size, server_name, server_port, silent=False, share=False):
|
181 |
-
if not silent:
|
182 |
-
print('Outputing stuff in', tmpdirname)
|
183 |
-
|
184 |
-
recon_fun = functools.partial(get_reconstructed_scene, tmpdirname, model, device, silent, image_size)
|
185 |
-
model_from_scene_fun = functools.partial(get_3D_model_from_scene, tmpdirname, silent)
|
186 |
-
with gradio.Blocks(css=""".gradio-container {margin: 0 !important; min-width: 100%};""", title="MASt3R Demo") as demo:
|
187 |
-
# scene state is save so that you can change conf_thr, cam_size... without rerunning the inference
|
188 |
-
scene = gradio.State(None)
|
189 |
-
gradio.HTML('<h2 style="text-align: center;">MASt3R Demo</h2>')
|
190 |
-
with gradio.Column():
|
191 |
-
inputfiles = gradio.File(file_count="multiple")
|
192 |
-
with gradio.Row():
|
193 |
-
with gradio.Column():
|
194 |
-
with gradio.Row():
|
195 |
-
lr1 = gradio.Slider(label="Coarse LR", value=0.07, minimum=0.01, maximum=0.2, step=0.01)
|
196 |
-
niter1 = gradio.Number(value=500, precision=0, minimum=0, maximum=10_000,
|
197 |
-
label="num_iterations", info="For coarse alignment!")
|
198 |
-
lr2 = gradio.Slider(label="Fine LR", value=0.014, minimum=0.005, maximum=0.05, step=0.001)
|
199 |
-
niter2 = gradio.Number(value=200, precision=0, minimum=0, maximum=100_000,
|
200 |
-
label="num_iterations", info="For refinement!")
|
201 |
-
optim_level = gradio.Dropdown(["coarse", "refine", "refine+depth"],
|
202 |
-
value='refine', label="OptLevel",
|
203 |
-
info="Optimization level")
|
204 |
-
with gradio.Row():
|
205 |
-
matching_conf_thr = gradio.Slider(label="Matching Confidence Thr", value=5.,
|
206 |
-
minimum=0., maximum=30., step=0.1,
|
207 |
-
info="Before Fallback to Regr3D!")
|
208 |
-
shared_intrinsics = gradio.Checkbox(value=False, label="Shared intrinsics",
|
209 |
-
info="Only optimize one set of intrinsics for all views")
|
210 |
-
scenegraph_type = gradio.Dropdown([("complete: all possible image pairs", "complete"),
|
211 |
-
("swin: sliding window", "swin"),
|
212 |
-
("logwin: sliding window with long range", "logwin"),
|
213 |
-
("oneref: match one image with all", "oneref")],
|
214 |
-
value='complete', label="Scenegraph",
|
215 |
-
info="Define how to make pairs",
|
216 |
-
interactive=True)
|
217 |
-
with gradio.Column(visible=False) as win_col:
|
218 |
-
winsize = gradio.Slider(label="Scene Graph: Window Size", value=1,
|
219 |
-
minimum=1, maximum=1, step=1)
|
220 |
-
win_cyclic = gradio.Checkbox(value=False, label="Cyclic sequence")
|
221 |
-
refid = gradio.Slider(label="Scene Graph: Id", value=0,
|
222 |
-
minimum=0, maximum=0, step=1, visible=False)
|
223 |
-
|
224 |
-
run_btn = gradio.Button("Run")
|
225 |
-
|
226 |
-
with gradio.Row():
|
227 |
-
# adjust the confidence threshold
|
228 |
-
min_conf_thr = gradio.Slider(label="min_conf_thr", value=1.5, minimum=0.0, maximum=10, step=0.1)
|
229 |
-
# adjust the camera size in the output pointcloud
|
230 |
-
cam_size = gradio.Slider(label="cam_size", value=0.2, minimum=0.001, maximum=1.0, step=0.001)
|
231 |
-
TSDF_thresh = gradio.Slider(label="TSDF Threshold", value=0., minimum=0., maximum=1., step=0.01)
|
232 |
-
with gradio.Row():
|
233 |
-
as_pointcloud = gradio.Checkbox(value=True, label="As pointcloud")
|
234 |
-
# two post process implemented
|
235 |
-
mask_sky = gradio.Checkbox(value=False, label="Mask sky")
|
236 |
-
clean_depth = gradio.Checkbox(value=True, label="Clean-up depthmaps")
|
237 |
-
transparent_cams = gradio.Checkbox(value=False, label="Transparent cameras")
|
238 |
-
|
239 |
-
outmodel = gradio.Model3D()
|
240 |
-
|
241 |
-
# events
|
242 |
-
scenegraph_type.change(set_scenegraph_options,
|
243 |
-
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
244 |
-
outputs=[win_col, winsize, win_cyclic, refid])
|
245 |
-
inputfiles.change(set_scenegraph_options,
|
246 |
-
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
247 |
-
outputs=[win_col, winsize, win_cyclic, refid])
|
248 |
-
win_cyclic.change(set_scenegraph_options,
|
249 |
-
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
250 |
-
outputs=[win_col, winsize, win_cyclic, refid])
|
251 |
-
run_btn.click(fn=recon_fun,
|
252 |
-
inputs=[inputfiles, optim_level, lr1, niter1, lr2, niter2, min_conf_thr, matching_conf_thr,
|
253 |
-
as_pointcloud, mask_sky, clean_depth, transparent_cams, cam_size,
|
254 |
-
scenegraph_type, winsize, win_cyclic, refid, TSDF_thresh, shared_intrinsics],
|
255 |
-
outputs=[scene, outmodel])
|
256 |
-
min_conf_thr.release(fn=model_from_scene_fun,
|
257 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
258 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
259 |
-
outputs=outmodel)
|
260 |
-
cam_size.change(fn=model_from_scene_fun,
|
261 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
262 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
263 |
-
outputs=outmodel)
|
264 |
-
TSDF_thresh.change(fn=model_from_scene_fun,
|
265 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
266 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
267 |
-
outputs=outmodel)
|
268 |
-
as_pointcloud.change(fn=model_from_scene_fun,
|
269 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
270 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
271 |
-
outputs=outmodel)
|
272 |
-
mask_sky.change(fn=model_from_scene_fun,
|
273 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
274 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
275 |
-
outputs=outmodel)
|
276 |
-
clean_depth.change(fn=model_from_scene_fun,
|
277 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
278 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
279 |
-
outputs=outmodel)
|
280 |
-
transparent_cams.change(model_from_scene_fun,
|
281 |
-
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
282 |
-
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
283 |
-
outputs=outmodel)
|
284 |
-
demo.launch(share=False, server_name=server_name, server_port=server_port)
|
285 |
-
|
286 |
|
287 |
if __name__ == '__main__':
|
288 |
parser = get_args_parser()
|
|
|
3 |
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
|
4 |
#
|
5 |
# --------------------------------------------------------
|
6 |
+
# gradio demo executable
|
7 |
# --------------------------------------------------------
|
|
|
|
|
8 |
import os
|
9 |
import torch
|
|
|
10 |
import tempfile
|
|
|
|
|
|
|
|
|
11 |
|
12 |
+
from mast3r.demo import get_args_parser, main_demo
|
|
|
13 |
|
14 |
from mast3r.model import AsymmetricMASt3R
|
15 |
from mast3r.utils.misc import hash_md5
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
import matplotlib.pyplot as pl
|
18 |
pl.ion()
|
19 |
|
20 |
torch.backends.cuda.matmul.allow_tf32 = True # for gpu >= Ampere and pytorch >= 1.12
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
if __name__ == '__main__':
|
23 |
parser = get_args_parser()
|
dust3r
CHANGED
@@ -1 +1 @@
|
|
1 |
-
Subproject commit
|
|
|
1 |
+
Subproject commit d99800a2d1d33f000c6f0d1c307dfb5a7a34fd53
|
mast3r/demo.py
ADDED
@@ -0,0 +1,277 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# Copyright (C) 2024-present Naver Corporation. All rights reserved.
|
3 |
+
# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
|
4 |
+
#
|
5 |
+
# --------------------------------------------------------
|
6 |
+
# sparse gradio demo functions
|
7 |
+
# --------------------------------------------------------
|
8 |
+
import math
|
9 |
+
import gradio
|
10 |
+
import os
|
11 |
+
import numpy as np
|
12 |
+
import functools
|
13 |
+
import trimesh
|
14 |
+
import copy
|
15 |
+
from scipy.spatial.transform import Rotation
|
16 |
+
|
17 |
+
from mast3r.cloud_opt.sparse_ga import sparse_global_alignment
|
18 |
+
from mast3r.cloud_opt.tsdf_optimizer import TSDFPostProcess
|
19 |
+
|
20 |
+
import mast3r.utils.path_to_dust3r # noqa
|
21 |
+
from dust3r.image_pairs import make_pairs
|
22 |
+
from dust3r.utils.image import load_images
|
23 |
+
from dust3r.utils.device import to_numpy
|
24 |
+
from dust3r.viz import add_scene_cam, CAM_COLORS, OPENGL, pts3d_to_trimesh, cat_meshes
|
25 |
+
from dust3r.demo import get_args_parser as dust3r_get_args_parser
|
26 |
+
|
27 |
+
import matplotlib.pyplot as pl
|
28 |
+
|
29 |
+
|
30 |
+
def get_args_parser():
|
31 |
+
parser = dust3r_get_args_parser()
|
32 |
+
parser.add_argument('--share', action='store_true')
|
33 |
+
|
34 |
+
actions = parser._actions
|
35 |
+
for action in actions:
|
36 |
+
if action.dest == 'model_name':
|
37 |
+
action.choices = ["MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"]
|
38 |
+
# change defaults
|
39 |
+
parser.prog = 'mast3r demo'
|
40 |
+
return parser
|
41 |
+
|
42 |
+
|
43 |
+
def _convert_scene_output_to_glb(outdir, imgs, pts3d, mask, focals, cams2world, cam_size=0.05,
|
44 |
+
cam_color=None, as_pointcloud=False,
|
45 |
+
transparent_cams=False, silent=False):
|
46 |
+
assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals)
|
47 |
+
pts3d = to_numpy(pts3d)
|
48 |
+
imgs = to_numpy(imgs)
|
49 |
+
focals = to_numpy(focals)
|
50 |
+
cams2world = to_numpy(cams2world)
|
51 |
+
|
52 |
+
scene = trimesh.Scene()
|
53 |
+
|
54 |
+
# full pointcloud
|
55 |
+
if as_pointcloud:
|
56 |
+
pts = np.concatenate([p[m.ravel()] for p, m in zip(pts3d, mask)])
|
57 |
+
col = np.concatenate([p[m] for p, m in zip(imgs, mask)])
|
58 |
+
pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))
|
59 |
+
scene.add_geometry(pct)
|
60 |
+
else:
|
61 |
+
meshes = []
|
62 |
+
for i in range(len(imgs)):
|
63 |
+
meshes.append(pts3d_to_trimesh(imgs[i], pts3d[i].reshape(imgs[i].shape), mask[i]))
|
64 |
+
mesh = trimesh.Trimesh(**cat_meshes(meshes))
|
65 |
+
scene.add_geometry(mesh)
|
66 |
+
|
67 |
+
# add each camera
|
68 |
+
for i, pose_c2w in enumerate(cams2world):
|
69 |
+
if isinstance(cam_color, list):
|
70 |
+
camera_edge_color = cam_color[i]
|
71 |
+
else:
|
72 |
+
camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)]
|
73 |
+
add_scene_cam(scene, pose_c2w, camera_edge_color,
|
74 |
+
None if transparent_cams else imgs[i], focals[i],
|
75 |
+
imsize=imgs[i].shape[1::-1], screen_width=cam_size)
|
76 |
+
|
77 |
+
rot = np.eye(4)
|
78 |
+
rot[:3, :3] = Rotation.from_euler('y', np.deg2rad(180)).as_matrix()
|
79 |
+
scene.apply_transform(np.linalg.inv(cams2world[0] @ OPENGL @ rot))
|
80 |
+
outfile = os.path.join(outdir, 'scene.glb')
|
81 |
+
if not silent:
|
82 |
+
print('(exporting 3D scene to', outfile, ')')
|
83 |
+
scene.export(file_obj=outfile)
|
84 |
+
return outfile
|
85 |
+
|
86 |
+
|
87 |
+
def get_3D_model_from_scene(outdir, silent, scene, min_conf_thr=2, as_pointcloud=False, mask_sky=False,
|
88 |
+
clean_depth=False, transparent_cams=False, cam_size=0.05, TSDF_thresh=0):
|
89 |
+
"""
|
90 |
+
extract 3D_model (glb file) from a reconstructed scene
|
91 |
+
"""
|
92 |
+
if scene is None:
|
93 |
+
return None
|
94 |
+
|
95 |
+
# get optimized values from scene
|
96 |
+
rgbimg = scene.imgs
|
97 |
+
focals = scene.get_focals().cpu()
|
98 |
+
cams2world = scene.get_im_poses().cpu()
|
99 |
+
|
100 |
+
# 3D pointcloud from depthmap, poses and intrinsics
|
101 |
+
if TSDF_thresh > 0:
|
102 |
+
tsdf = TSDFPostProcess(scene, TSDF_thresh=TSDF_thresh)
|
103 |
+
pts3d, _, confs = to_numpy(tsdf.get_dense_pts3d(clean_depth=clean_depth))
|
104 |
+
else:
|
105 |
+
pts3d, _, confs = to_numpy(scene.get_dense_pts3d(clean_depth=clean_depth))
|
106 |
+
msk = to_numpy([c > min_conf_thr for c in confs])
|
107 |
+
return _convert_scene_output_to_glb(outdir, rgbimg, pts3d, msk, focals, cams2world, as_pointcloud=as_pointcloud,
|
108 |
+
transparent_cams=transparent_cams, cam_size=cam_size, silent=silent)
|
109 |
+
|
110 |
+
|
111 |
+
def get_reconstructed_scene(outdir, model, device, silent, image_size, filelist, optim_level, lr1, niter1, lr2, niter2,
|
112 |
+
min_conf_thr, matching_conf_thr, as_pointcloud, mask_sky, clean_depth, transparent_cams,
|
113 |
+
cam_size, scenegraph_type, winsize, win_cyclic, refid, TSDF_thresh, shared_intrinsics,
|
114 |
+
**kw):
|
115 |
+
"""
|
116 |
+
from a list of images, run mast3r inference, sparse global aligner.
|
117 |
+
then run get_3D_model_from_scene
|
118 |
+
"""
|
119 |
+
imgs = load_images(filelist, size=image_size, verbose=not silent)
|
120 |
+
if len(imgs) == 1:
|
121 |
+
imgs = [imgs[0], copy.deepcopy(imgs[0])]
|
122 |
+
imgs[1]['idx'] = 1
|
123 |
+
filelist = [filelist[0], filelist[0] + '_2']
|
124 |
+
|
125 |
+
scene_graph_params = [scenegraph_type]
|
126 |
+
if scenegraph_type in ["swin", "logwin"]:
|
127 |
+
scene_graph_params.append(str(winsize))
|
128 |
+
elif scenegraph_type == "oneref":
|
129 |
+
scene_graph_params.append(str(refid))
|
130 |
+
if scenegraph_type in ["swin", "logwin"] and not win_cyclic:
|
131 |
+
scene_graph_params.append('noncyclic')
|
132 |
+
scene_graph = '-'.join(scene_graph_params)
|
133 |
+
pairs = make_pairs(imgs, scene_graph=scene_graph, prefilter=None, symmetrize=True)
|
134 |
+
if optim_level == 'coarse':
|
135 |
+
niter2 = 0
|
136 |
+
# Sparse GA (forward mast3r -> matching -> 3D optim -> 2D refinement -> triangulation)
|
137 |
+
scene = sparse_global_alignment(filelist, pairs, os.path.join(outdir, 'cache'),
|
138 |
+
model, lr1=lr1, niter1=niter1, lr2=lr2, niter2=niter2, device=device,
|
139 |
+
opt_depth='depth' in optim_level, shared_intrinsics=shared_intrinsics,
|
140 |
+
matching_conf_thr=matching_conf_thr, **kw)
|
141 |
+
outfile = get_3D_model_from_scene(outdir, silent, scene, min_conf_thr, as_pointcloud, mask_sky,
|
142 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh)
|
143 |
+
return scene, outfile
|
144 |
+
|
145 |
+
|
146 |
+
def set_scenegraph_options(inputfiles, win_cyclic, refid, scenegraph_type):
|
147 |
+
num_files = len(inputfiles) if inputfiles is not None else 1
|
148 |
+
show_win_controls = scenegraph_type in ["swin", "logwin"]
|
149 |
+
show_winsize = scenegraph_type in ["swin", "logwin"]
|
150 |
+
show_cyclic = scenegraph_type in ["swin", "logwin"]
|
151 |
+
max_winsize, min_winsize = 1, 1
|
152 |
+
if scenegraph_type == "swin":
|
153 |
+
if win_cyclic:
|
154 |
+
max_winsize = max(1, math.ceil((num_files - 1) / 2))
|
155 |
+
else:
|
156 |
+
max_winsize = num_files - 1
|
157 |
+
elif scenegraph_type == "logwin":
|
158 |
+
if win_cyclic:
|
159 |
+
half_size = math.ceil((num_files - 1) / 2)
|
160 |
+
max_winsize = max(1, math.ceil(math.log(half_size, 2)))
|
161 |
+
else:
|
162 |
+
max_winsize = max(1, math.ceil(math.log(num_files, 2)))
|
163 |
+
winsize = gradio.Slider(label="Scene Graph: Window Size", value=max_winsize,
|
164 |
+
minimum=min_winsize, maximum=max_winsize, step=1, visible=show_winsize)
|
165 |
+
win_cyclic = gradio.Checkbox(value=win_cyclic, label="Cyclic sequence", visible=show_cyclic)
|
166 |
+
win_col = gradio.Column(visible=show_win_controls)
|
167 |
+
refid = gradio.Slider(label="Scene Graph: Id", value=0, minimum=0,
|
168 |
+
maximum=num_files - 1, step=1, visible=scenegraph_type == 'oneref')
|
169 |
+
return win_col, winsize, win_cyclic, refid
|
170 |
+
|
171 |
+
|
172 |
+
def main_demo(tmpdirname, model, device, image_size, server_name, server_port, silent=False, share=False):
|
173 |
+
if not silent:
|
174 |
+
print('Outputing stuff in', tmpdirname)
|
175 |
+
|
176 |
+
recon_fun = functools.partial(get_reconstructed_scene, tmpdirname, model, device, silent, image_size)
|
177 |
+
model_from_scene_fun = functools.partial(get_3D_model_from_scene, tmpdirname, silent)
|
178 |
+
with gradio.Blocks(css=""".gradio-container {margin: 0 !important; min-width: 100%};""", title="MASt3R Demo") as demo:
|
179 |
+
# scene state is save so that you can change conf_thr, cam_size... without rerunning the inference
|
180 |
+
scene = gradio.State(None)
|
181 |
+
gradio.HTML('<h2 style="text-align: center;">MASt3R Demo</h2>')
|
182 |
+
with gradio.Column():
|
183 |
+
inputfiles = gradio.File(file_count="multiple")
|
184 |
+
with gradio.Row():
|
185 |
+
with gradio.Column():
|
186 |
+
with gradio.Row():
|
187 |
+
lr1 = gradio.Slider(label="Coarse LR", value=0.07, minimum=0.01, maximum=0.2, step=0.01)
|
188 |
+
niter1 = gradio.Number(value=500, precision=0, minimum=0, maximum=10_000,
|
189 |
+
label="num_iterations", info="For coarse alignment!")
|
190 |
+
lr2 = gradio.Slider(label="Fine LR", value=0.014, minimum=0.005, maximum=0.05, step=0.001)
|
191 |
+
niter2 = gradio.Number(value=200, precision=0, minimum=0, maximum=100_000,
|
192 |
+
label="num_iterations", info="For refinement!")
|
193 |
+
optim_level = gradio.Dropdown(["coarse", "refine", "refine+depth"],
|
194 |
+
value='refine', label="OptLevel",
|
195 |
+
info="Optimization level")
|
196 |
+
with gradio.Row():
|
197 |
+
matching_conf_thr = gradio.Slider(label="Matching Confidence Thr", value=5.,
|
198 |
+
minimum=0., maximum=30., step=0.1,
|
199 |
+
info="Before Fallback to Regr3D!")
|
200 |
+
shared_intrinsics = gradio.Checkbox(value=False, label="Shared intrinsics",
|
201 |
+
info="Only optimize one set of intrinsics for all views")
|
202 |
+
scenegraph_type = gradio.Dropdown([("complete: all possible image pairs", "complete"),
|
203 |
+
("swin: sliding window", "swin"),
|
204 |
+
("logwin: sliding window with long range", "logwin"),
|
205 |
+
("oneref: match one image with all", "oneref")],
|
206 |
+
value='complete', label="Scenegraph",
|
207 |
+
info="Define how to make pairs",
|
208 |
+
interactive=True)
|
209 |
+
with gradio.Column(visible=False) as win_col:
|
210 |
+
winsize = gradio.Slider(label="Scene Graph: Window Size", value=1,
|
211 |
+
minimum=1, maximum=1, step=1)
|
212 |
+
win_cyclic = gradio.Checkbox(value=False, label="Cyclic sequence")
|
213 |
+
refid = gradio.Slider(label="Scene Graph: Id", value=0,
|
214 |
+
minimum=0, maximum=0, step=1, visible=False)
|
215 |
+
|
216 |
+
run_btn = gradio.Button("Run")
|
217 |
+
|
218 |
+
with gradio.Row():
|
219 |
+
# adjust the confidence threshold
|
220 |
+
min_conf_thr = gradio.Slider(label="min_conf_thr", value=1.5, minimum=0.0, maximum=10, step=0.1)
|
221 |
+
# adjust the camera size in the output pointcloud
|
222 |
+
cam_size = gradio.Slider(label="cam_size", value=0.2, minimum=0.001, maximum=1.0, step=0.001)
|
223 |
+
TSDF_thresh = gradio.Slider(label="TSDF Threshold", value=0., minimum=0., maximum=1., step=0.01)
|
224 |
+
with gradio.Row():
|
225 |
+
as_pointcloud = gradio.Checkbox(value=True, label="As pointcloud")
|
226 |
+
# two post process implemented
|
227 |
+
mask_sky = gradio.Checkbox(value=False, label="Mask sky")
|
228 |
+
clean_depth = gradio.Checkbox(value=True, label="Clean-up depthmaps")
|
229 |
+
transparent_cams = gradio.Checkbox(value=False, label="Transparent cameras")
|
230 |
+
|
231 |
+
outmodel = gradio.Model3D()
|
232 |
+
|
233 |
+
# events
|
234 |
+
scenegraph_type.change(set_scenegraph_options,
|
235 |
+
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
236 |
+
outputs=[win_col, winsize, win_cyclic, refid])
|
237 |
+
inputfiles.change(set_scenegraph_options,
|
238 |
+
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
239 |
+
outputs=[win_col, winsize, win_cyclic, refid])
|
240 |
+
win_cyclic.change(set_scenegraph_options,
|
241 |
+
inputs=[inputfiles, win_cyclic, refid, scenegraph_type],
|
242 |
+
outputs=[win_col, winsize, win_cyclic, refid])
|
243 |
+
run_btn.click(fn=recon_fun,
|
244 |
+
inputs=[inputfiles, optim_level, lr1, niter1, lr2, niter2, min_conf_thr, matching_conf_thr,
|
245 |
+
as_pointcloud, mask_sky, clean_depth, transparent_cams, cam_size,
|
246 |
+
scenegraph_type, winsize, win_cyclic, refid, TSDF_thresh, shared_intrinsics],
|
247 |
+
outputs=[scene, outmodel])
|
248 |
+
min_conf_thr.release(fn=model_from_scene_fun,
|
249 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
250 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
251 |
+
outputs=outmodel)
|
252 |
+
cam_size.change(fn=model_from_scene_fun,
|
253 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
254 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
255 |
+
outputs=outmodel)
|
256 |
+
TSDF_thresh.change(fn=model_from_scene_fun,
|
257 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
258 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
259 |
+
outputs=outmodel)
|
260 |
+
as_pointcloud.change(fn=model_from_scene_fun,
|
261 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
262 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
263 |
+
outputs=outmodel)
|
264 |
+
mask_sky.change(fn=model_from_scene_fun,
|
265 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
266 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
267 |
+
outputs=outmodel)
|
268 |
+
clean_depth.change(fn=model_from_scene_fun,
|
269 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
270 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
271 |
+
outputs=outmodel)
|
272 |
+
transparent_cams.change(model_from_scene_fun,
|
273 |
+
inputs=[scene, min_conf_thr, as_pointcloud, mask_sky,
|
274 |
+
clean_depth, transparent_cams, cam_size, TSDF_thresh],
|
275 |
+
outputs=outmodel)
|
276 |
+
demo.launch(share=share, server_name=server_name, server_port=server_port)
|
277 |
+
|