-
Notifications
You must be signed in to change notification settings - Fork 3
/
mapping.py
382 lines (330 loc) · 15.2 KB
/
mapping.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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
import argparse
import os
import shutil
import numpy as np
import pycolmap
import tempfile
import cv2
import open3d as o3d
from pathlib import Path
from autolabel.utils import Scene, transform_points, Camera
from autolabel.undistort import ImageUndistorter
from hloc import (extract_features, match_features, reconstruction,
pairs_from_exhaustive, pairs_from_retrieval)
from hloc.utils import viz_3d
def read_args():
parser = argparse.ArgumentParser()
parser.add_argument('scene', help="Scene to infer poses for.")
parser.add_argument('--debug', action='store_true')
parser.add_argument('--vis', action='store_true')
return parser.parse_args()
class HLoc:
def __init__(self, tmp_dir, scene, flags):
self.flags = flags
self.scene = scene
self.scene_path = Path(self.scene.path)
self.exhaustive = len((self.scene.raw_rgb_paths())) < 250
self.tmp_dir = Path(tmp_dir)
self.sfm_pairs = self.tmp_dir / 'sfm-pairs.txt'
self.loc_pairs = self.tmp_dir / 'sfm-pairs-loc.txt'
self.features = self.tmp_dir / 'features.h5'
self.matches = self.tmp_dir / 'matches.h5'
self.feature_conf = extract_features.confs['superpoint_aachen']
self.retrieval_conf = extract_features.confs['netvlad']
self.matcher_conf = match_features.confs['superglue']
def _run_sfm(self):
image_dir = Path(self.scene.path) / 'raw_rgb'
image_list = []
image_paths = self.scene.raw_rgb_paths()
image_list_path = []
indices = np.arange(len(image_paths))
for index in indices:
image_list.append(image_paths[index])
image_list_path.append(
str(Path(image_paths[index]).relative_to(image_dir)))
if self.exhaustive:
extract_features.main(self.feature_conf,
image_dir,
feature_path=self.features,
image_list=image_list_path)
pairs_from_exhaustive.main(self.sfm_pairs,
image_list=image_list_path)
match_features.main(self.matcher_conf,
self.sfm_pairs,
features=self.features,
matches=self.matches)
model = reconstruction.main(
self.tmp_dir,
image_dir,
self.sfm_pairs,
self.features,
self.matches,
image_list=image_list_path,
camera_mode=pycolmap.CameraMode.SINGLE,
image_options={'camera_model': "OPENCV"},
mapper_options={
'ba_refine_principal_point': True,
'ba_refine_extra_params': True,
'ba_refine_focal_length': True
})
else:
retrieval_path = extract_features.main(self.retrieval_conf,
image_dir,
self.tmp_dir,
image_list=image_list_path)
pairs_from_retrieval.main(retrieval_path,
self.sfm_pairs,
num_matched=50)
feature_path = extract_features.main(self.feature_conf,
image_dir,
self.tmp_dir,
image_list=image_list_path)
match_path = match_features.main(self.matcher_conf,
self.sfm_pairs,
self.feature_conf['output'],
self.tmp_dir,
matches=self.matches)
model = reconstruction.main(
self.tmp_dir,
image_dir,
self.sfm_pairs,
feature_path,
match_path,
image_list=image_list_path,
camera_mode=pycolmap.CameraMode.SINGLE,
image_options={'camera_model': "OPENCV"},
mapper_options={
'ba_refine_principal_point': True,
'ba_refine_extra_params': True,
'ba_refine_focal_length': True
})
if self.flags.vis:
fig = viz_3d.init_figure()
viz_3d.plot_reconstruction(fig,
model,
color='rgba(255,0,0,0.5)',
name="mapping")
fig.show()
if self.flags.debug:
# Save mapping metadata if running in debug mode.
colmap_output_dir = os.path.join(self.scene.path, 'colmap_output')
os.makedirs(colmap_output_dir, exist_ok=True)
model.write_text(colmap_output_dir)
# Save the intrinsics matrix and the distortion parameters.
assert (len(model.cameras) == 1 and 1 in model.cameras)
(focal_length_x, focal_length_y, c_x, c_y, k_1, k_2, p_1,
p_2) = model.cameras[1].params
self.colmap_K = np.eye(3)
self.colmap_K[0, 0] = focal_length_x
self.colmap_K[1, 1] = focal_length_y
self.colmap_K[0, 2] = c_x
self.colmap_K[1, 2] = c_y
self.colmap_distortion_params = np.array([k_1, k_2, p_1, p_2])
np.savetxt(fname=os.path.join(self.scene.path, 'intrinsics.txt'),
X=self.colmap_K)
np.savetxt(fname=os.path.join(self.scene.path,
'distortion_parameters.txt'),
X=self.colmap_distortion_params)
def _undistort_images(self):
print("Undistorting images according to the estimated intrinsics...")
undistorted_image_folder = os.path.join(self.scene.path, "rgb")
undistorted_depth_folder = os.path.join(self.scene.path, "depth")
os.makedirs(undistorted_image_folder, exist_ok=True)
os.makedirs(undistorted_depth_folder, exist_ok=True)
color_undistorter = ImageUndistorter(K=self.colmap_K,
D=self.colmap_distortion_params,
H=self.scene.camera.size[1],
W=self.scene.camera.size[0])
depth_camera = Camera(self.colmap_K, self.scene.camera.size).scale(
self.scene.depth_size())
depth_undistorter = ImageUndistorter(K=depth_camera.camera_matrix,
D=self.colmap_distortion_params,
H=depth_camera.size[1],
W=depth_camera.size[0])
# Undistort all the images and save the undistorted versions.
image_paths = self.scene.raw_rgb_paths()
for image_path in image_paths:
image = cv2.imread(image_path, cv2.IMREAD_UNCHANGED)
undistorted_image = color_undistorter.undistort_image(image=image)
cv2.imwrite(img=undistorted_image,
filename=os.path.join(undistorted_image_folder,
os.path.basename(image_path)))
depth_paths = self.scene.raw_depth_paths()
for depth_path in depth_paths:
depth = cv2.imread(depth_path, cv2.IMREAD_UNCHANGED)
undistorted_depth = depth_undistorter.undistort_image(image=depth)
cv2.imwrite(img=undistorted_depth,
filename=os.path.join(undistorted_depth_folder,
os.path.basename(depth_path)))
def run(self):
self._run_sfm()
self._undistort_images()
class ScaleEstimation:
min_depth = 0.05
def __init__(self, scene, colmap_dir):
self.scene = scene
self.colmap_dir = colmap_dir
self.reconstruction = pycolmap.Reconstruction(colmap_dir)
self._read_trajectory()
self._read_depth_maps()
def _read_depth_maps(self):
self.depth_maps = {}
for path in self.scene.depth_paths():
frame_name = os.path.basename(path).split('.')[0]
self.depth_maps[frame_name] = cv2.imread(path, -1) / 1000.0
depth_shape = next(iter(self.depth_maps.values())).shape
depth_size = np.array([depth_shape[1], depth_shape[0]],
dtype=np.float64)
self.depth_to_color_ratio = depth_size / np.array(
self.scene.camera.size, dtype=np.float64)
def _read_trajectory(self):
poses = []
for image in self.reconstruction.images.values():
T_CW = np.eye(4)
T_CW[:3, :3] = image.rotmat()
T_CW[:3, 3] = image.tvec
frame_name = image.name.split('.')[0]
poses.append((frame_name, T_CW))
self.poses = dict(poses)
def _lookup_depth(self, frame, xy):
xy_depth = np.floor(self.depth_to_color_ratio * xy).astype(int)
return self.depth_maps[frame][xy_depth[1], xy_depth[0]]
def _estimate_scale(self):
images = self.reconstruction.images
point_depths = []
measured_depths = []
for image in images.values():
frame_name = image.name.split('.')[0]
points = image.get_valid_points2D()
points3D = self.reconstruction.points3D
for point in points:
depth_map_value = self._lookup_depth(frame_name, point.xy)
if depth_map_value < self.min_depth:
continue
T_CW = self.poses[frame_name]
point3D = points3D[point.point3D_id]
p_C = transform_points(T_CW, point3D.xyz)
measured_depths.append(depth_map_value)
point_depths.append(p_C[2])
point_depths = np.stack(point_depths)
measured_depths = np.stack(measured_depths)
scales = measured_depths / point_depths
return self._ransac(scales)
def _ransac(self, scales):
best_set = None
best_inlier_count = 0
indices = np.arange(0, scales.shape[0])
inlier_threshold = np.median(scales) * 1e-2
for i in range(10000):
selected = np.random.choice(indices)
estimate = scales[selected]
inliers = np.abs(scales - estimate) < inlier_threshold
inlier_count = inliers.sum()
if inlier_count > best_inlier_count:
best_set = scales[inliers]
best_inlier_count = inlier_count
print(
f"Scale estimation inlier count: {best_inlier_count} / {scales.size}"
)
return np.mean(best_set)
def _scale_poses(self, ratio):
scaled_poses = {}
for key, pose in self.poses.items():
new_pose = pose.copy()
new_pose[:3, 3] *= ratio
scaled_poses[key] = new_pose
return scaled_poses
def run(self):
scale_ratio = self._estimate_scale()
return self._scale_poses(scale_ratio)
class PoseSaver:
def __init__(self, scene, scaled_poses):
self.scene = scene
self.poses = scaled_poses
def compute_bbox(self, poses):
"""
poses: Metrically scaled transforms from camera to world frame.
"""
# Compute axis-aligned bounding box of the depth values in world frame.
# Then get the center.
min_bounds = np.zeros(3)
max_bounds = np.zeros(3)
depth_frame = o3d.io.read_image(self.scene.depth_paths()[0])
depth_size = np.asarray(depth_frame).shape[::-1]
K = self.scene.camera.scale(depth_size).camera_matrix
intrinsics = o3d.camera.PinholeCameraIntrinsic(int(depth_size[0]),
int(depth_size[1]),
K[0, 0], K[1, 1],
K[0, 2], K[1, 2])
pc = o3d.geometry.PointCloud()
depth_frames = dict([(os.path.basename(p).split('.')[0], p)
for p in self.scene.depth_paths()])
items = [item for item in poses.items()]
stride = max(len(self.scene.depth_paths()) // 100, 1)
for key, T_WC in items[::stride]:
if key not in depth_frames:
print("WARNING: Can't find depth image {key}.png")
continue
depth = o3d.io.read_image(f"{depth_frames[key]}")
pc_C = o3d.geometry.PointCloud.create_from_depth_image(
depth, depth_scale=1000.0, intrinsic=intrinsics)
pc_C = np.asarray(pc_C.points)
pc_W = transform_points(T_WC, pc_C)
min_bounds = np.minimum(min_bounds, pc_W.min(axis=0))
max_bounds = np.maximum(max_bounds, pc_W.max(axis=0))
pc += o3d.geometry.PointCloud(
o3d.utility.Vector3dVector(pc_W)).uniform_down_sample(50)
filtered, _ = pc.remove_statistical_outlier(nb_neighbors=20,
std_ratio=2.0)
bbox = filtered.get_oriented_bounding_box(robust=True)
T = np.eye(4)
T[:3, :3] = bbox.R.T
o3d_aabb = o3d.geometry.PointCloud(filtered).transform(
T).get_axis_aligned_bounding_box()
center = o3d_aabb.get_center()
T[:3, 3] = -center
aabb = np.zeros((2, 3))
aabb[0, :] = o3d_aabb.get_min_bound() - center
aabb[1, :] = o3d_aabb.get_max_bound() - center
return T, aabb, filtered
def _write_poses(self, poses):
pose_dir = os.path.join(self.scene.path, 'pose')
os.makedirs(pose_dir, exist_ok=True)
for key, T_CW in poses.items():
pose_file = os.path.join(pose_dir, f'{key}.txt')
np.savetxt(pose_file, T_CW)
def _write_bounds(self, bounds):
with open(os.path.join(self.scene.path, 'bbox.txt'), 'wt') as f:
min_str = " ".join([str(x) for x in bounds[0]])
max_str = " ".join([str(x) for x in bounds[1]])
f.write(f"{min_str} {max_str} 0.01")
def run(self):
T_WCs = {}
for key, T_CW in self.poses.items():
T_WCs[key] = np.linalg.inv(T_CW)
T, aabb, point_cloud = self.compute_bbox(T_WCs)
T_CWs = {}
for key, T_WC in T_WCs.items():
T_CWs[key] = np.linalg.inv(T @ T_WC)
self._write_poses(T_CWs)
self._write_bounds(aabb)
class Pipeline:
def __init__(self, flags):
self.tmp_dir = tempfile.mkdtemp()
self.flags = flags
self.scene = Scene(flags.scene)
def run(self):
hloc = HLoc(self.tmp_dir, self.scene, self.flags)
hloc.run()
# Camera intrinsics might have changed so reload the scene.
self.scene = Scene(self.scene.path)
scale_estimation = ScaleEstimation(self.scene, self.tmp_dir)
scaled_poses = scale_estimation.run()
pose_saver = PoseSaver(self.scene, scaled_poses)
pose_saver.run()
if self.flags.debug:
shutil.move(str(self.tmp_dir), "/tmp/sfm_debug")
else:
shutil.rmtree(self.tmp_dir)
if __name__ == "__main__":
Pipeline(read_args()).run()