forked from NVlabs/FoundationPose
-
Notifications
You must be signed in to change notification settings - Fork 0
/
estimater.py
276 lines (217 loc) · 10.3 KB
/
estimater.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
# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
from Utils import *
from datareader import *
import itertools
from learning.training.predict_score import *
from learning.training.predict_pose_refine import *
import yaml
class FoundationPose:
def __init__(self, model_pts, model_normals, symmetry_tfs=None, mesh=None, scorer:ScorePredictor=None, refiner:PoseRefinePredictor=None, glctx=None, debug=0, debug_dir='/home/bowen/debug/novel_pose_debug/'):
self.gt_pose = None
self.ignore_normal_flip = True
self.debug = debug
self.debug_dir = debug_dir
os.makedirs(debug_dir, exist_ok=True)
self.reset_object(model_pts, model_normals, symmetry_tfs=symmetry_tfs, mesh=mesh)
self.make_rotation_grid(min_n_views=40, inplane_step=60)
self.glctx = glctx
if scorer is not None:
self.scorer = scorer
else:
self.scorer = ScorePredictor()
if refiner is not None:
self.refiner = refiner
else:
self.refiner = PoseRefinePredictor()
self.pose_last = None # Used for tracking; per the centered mesh
def reset_object(self, model_pts, model_normals, symmetry_tfs=None, mesh=None):
max_xyz = mesh.vertices.max(axis=0)
min_xyz = mesh.vertices.min(axis=0)
self.model_center = (min_xyz+max_xyz)/2
if mesh is not None:
self.mesh_ori = mesh.copy()
mesh = mesh.copy()
mesh.vertices = mesh.vertices - self.model_center.reshape(1,3)
model_pts = mesh.vertices
self.diameter = compute_mesh_diameter(model_pts=mesh.vertices, n_sample=10000)
self.vox_size = max(self.diameter/20.0, 0.003)
logging.info(f'self.diameter:{self.diameter}, vox_size:{self.vox_size}')
self.dist_bin = self.vox_size/2
self.angle_bin = 20 # Deg
pcd = toOpen3dCloud(model_pts, normals=model_normals)
pcd = pcd.voxel_down_sample(self.vox_size)
self.max_xyz = np.asarray(pcd.points).max(axis=0)
self.min_xyz = np.asarray(pcd.points).min(axis=0)
self.pts = torch.tensor(np.asarray(pcd.points), dtype=torch.float32, device='cuda')
self.normals = F.normalize(torch.tensor(np.asarray(pcd.normals), dtype=torch.float32, device='cuda'), dim=-1)
logging.info(f'self.pts:{self.pts.shape}')
self.mesh_path = None
self.mesh = mesh
if self.mesh is not None:
self.mesh_path = f'/tmp/{uuid.uuid4()}.obj'
self.mesh.export(self.mesh_path)
self.mesh_tensors = make_mesh_tensors(self.mesh)
if symmetry_tfs is None:
self.symmetry_tfs = torch.eye(4).float().cuda()[None]
else:
self.symmetry_tfs = torch.as_tensor(symmetry_tfs, device='cuda', dtype=torch.float)
logging.info("reset done")
def get_tf_to_centered_mesh(self):
tf_to_center = torch.eye(4, dtype=torch.float, device='cuda')
tf_to_center[:3,3] = -torch.as_tensor(self.model_center, device='cuda', dtype=torch.float)
return tf_to_center
def to_device(self, s='cuda:0'):
for k in self.__dict__:
self.__dict__[k] = self.__dict__[k]
if torch.is_tensor(self.__dict__[k]) or isinstance(self.__dict__[k], nn.Module):
logging.info(f"Moving {k} to device {s}")
self.__dict__[k] = self.__dict__[k].to(s)
for k in self.mesh_tensors:
logging.info(f"Moving {k} to device {s}")
self.mesh_tensors[k] = self.mesh_tensors[k].to(s)
if self.refiner is not None:
self.refiner.model.to(s)
if self.scorer is not None:
self.scorer.model.to(s)
if self.glctx is not None:
self.glctx = dr.RasterizeCudaContext(s)
def make_rotation_grid(self, min_n_views=40, inplane_step=60):
cam_in_obs = sample_views_icosphere(n_views=min_n_views)
logging.info(f'cam_in_obs:{cam_in_obs.shape}')
rot_grid = []
for i in range(len(cam_in_obs)):
for inplane_rot in np.deg2rad(np.arange(0, 360, inplane_step)):
cam_in_ob = cam_in_obs[i]
R_inplane = euler_matrix(0,0,inplane_rot)
cam_in_ob = cam_in_ob@R_inplane
ob_in_cam = np.linalg.inv(cam_in_ob)
rot_grid.append(ob_in_cam)
rot_grid = np.asarray(rot_grid)
logging.info(f"rot_grid:{rot_grid.shape}")
rot_grid = mycpp.cluster_poses(30, 99999, rot_grid, self.symmetry_tfs.data.cpu().numpy())
rot_grid = np.asarray(rot_grid)
logging.info(f"after cluster, rot_grid:{rot_grid.shape}")
self.rot_grid = torch.as_tensor(rot_grid, device='cuda', dtype=torch.float)
logging.info(f"self.rot_grid: {self.rot_grid.shape}")
def generate_random_pose_hypo(self, K, rgb, depth, mask, scene_pts=None):
'''
@scene_pts: torch tensor (N,3)
'''
ob_in_cams = self.rot_grid.clone()
center = self.guess_translation(depth=depth, mask=mask, K=K)
ob_in_cams[:,:3,3] = torch.tensor(center, device='cuda', dtype=torch.float).reshape(1,3)
return ob_in_cams
def guess_translation(self, depth, mask, K):
vs,us = np.where(mask>0)
if len(us)==0:
logging.info(f'mask is all zero')
return np.zeros((3))
uc = (us.min()+us.max())/2.0
vc = (vs.min()+vs.max())/2.0
valid = mask.astype(bool) & (depth>=0.1)
if not valid.any():
logging.info(f"valid is empty")
return np.zeros((3))
zc = np.median(depth[valid])
#zc = np.mean(depth[valid])
center = (np.linalg.inv(K)@np.asarray([uc,vc,1]).reshape(3,1))*zc
if self.debug>=2:
pcd = toOpen3dCloud(center.reshape(1,3))
o3d.io.write_point_cloud(f'{self.debug_dir}/init_center.ply', pcd)
return center.reshape(3)
def register(self, K, rgb, depth, ob_mask, ob_id=None, glctx=None, iteration=5):
'''Copmute pose from given pts to self.pcd
@pts: (N,3) np array, downsampled scene points
'''
set_seed(0)
logging.info('Welcome')
if self.glctx is None:
if glctx is None:
self.glctx = dr.RasterizeCudaContext()
# self.glctx = dr.RasterizeGLContext()
else:
self.glctx = glctx
depth = erode_depth(depth, radius=2, device='cuda')
depth = bilateral_filter_depth(depth, radius=2, device='cuda')
if self.debug>=2:
xyz_map = depth2xyzmap(depth, K)
valid = xyz_map[...,2]>=0.1
pcd = toOpen3dCloud(xyz_map[valid], rgb[valid])
o3d.io.write_point_cloud(f'{self.debug_dir}/scene_raw.ply',pcd)
cv2.imwrite(f'{self.debug_dir}/ob_mask.png', (ob_mask*255.0).clip(0,255))
normal_map = None
valid = (depth>=0.1) & (ob_mask>0)
if valid.sum()<4:
logging.info(f'valid too small, return')
pose = np.eye(4)
pose[:3,3] = self.guess_translation(depth=depth, mask=ob_mask, K=K)
return pose
if self.debug>=2:
imageio.imwrite(f'{self.debug_dir}/color.png', rgb)
cv2.imwrite(f'{self.debug_dir}/depth.png', (depth*1000).astype(np.uint16))
valid = xyz_map[...,2]>=0.1
pcd = toOpen3dCloud(xyz_map[valid], rgb[valid])
o3d.io.write_point_cloud(f'{self.debug_dir}/scene_complete.ply',pcd)
self.H, self.W = depth.shape[:2]
self.K = K
self.ob_id = ob_id
self.ob_mask = ob_mask
poses = self.generate_random_pose_hypo(K=K, rgb=rgb, depth=depth, mask=ob_mask, scene_pts=None)
poses = poses.data.cpu().numpy()
logging.info(f'poses:{poses.shape}')
center = self.guess_translation(depth=depth, mask=ob_mask, K=K)
poses = torch.as_tensor(poses, device='cuda', dtype=torch.float)
poses[:,:3,3] = torch.as_tensor(center.reshape(1,3), device='cuda')
add_errs = self.compute_add_err_to_gt_pose(poses)
logging.info(f"after viewpoint, add_errs min:{add_errs.min()}")
xyz_map = depth2xyzmap(depth, K)
poses, vis = self.refiner.predict(mesh=self.mesh, mesh_tensors=self.mesh_tensors, rgb=rgb, depth=depth, K=K, ob_in_cams=poses.data.cpu().numpy(), normal_map=normal_map, xyz_map=xyz_map, glctx=self.glctx, mesh_diameter=self.diameter, iteration=iteration, get_vis=self.debug>=2)
if vis is not None:
imageio.imwrite(f'{self.debug_dir}/vis_refiner.png', vis)
scores, vis = self.scorer.predict(mesh=self.mesh, rgb=rgb, depth=depth, K=K, ob_in_cams=poses.data.cpu().numpy(), normal_map=normal_map, mesh_tensors=self.mesh_tensors, glctx=self.glctx, mesh_diameter=self.diameter, get_vis=self.debug>=2)
if vis is not None:
imageio.imwrite(f'{self.debug_dir}/vis_score.png', vis)
add_errs = self.compute_add_err_to_gt_pose(poses)
logging.info(f"final, add_errs min:{add_errs.min()}")
ids = torch.as_tensor(scores).argsort(descending=True)
logging.info(f'sort ids:{ids}')
scores = scores[ids]
poses = poses[ids]
logging.info(f'sorted scores:{scores}')
best_pose = poses[0]@self.get_tf_to_centered_mesh()
self.pose_last = poses[0]
self.best_id = ids[0]
self.poses = poses
self.scores = scores
return best_pose.data.cpu().numpy()
def compute_add_err_to_gt_pose(self, poses):
'''
@poses: wrt. the centered mesh
'''
return -torch.ones(len(poses), device='cuda', dtype=torch.float)
def track_one(self, rgb, depth, K, ob_mask, iteration, extra={}):
if self.pose_last is None:
logging.info("Please init pose by register first")
raise RuntimeError
logging.info("Welcome")
poses = self.pose_last.data.cpu().numpy().reshape(1,4,4)
if ob_mask is not None:
center = self.guess_translation(depth=depth, mask=ob_mask, K=K)
poses[0,:3,3] = (poses[0,:3,3] + center.reshape(3)) / 2.0
depth = torch.as_tensor(depth, device='cuda', dtype=torch.float)
depth = erode_depth(depth, radius=2, device='cuda')
depth = bilateral_filter_depth(depth, radius=2, device='cuda')
logging.info("depth processing done")
xyz_map = depth2xyzmap_batch(depth[None], torch.as_tensor(K, dtype=torch.float, device='cuda')[None], zfar=np.inf)[0]
pose, vis = self.refiner.predict(mesh=self.mesh, mesh_tensors=self.mesh_tensors, rgb=rgb, depth=depth, K=K, ob_in_cams=poses, normal_map=None, xyz_map=xyz_map, mesh_diameter=self.diameter, glctx=self.glctx, iteration=iteration, get_vis=self.debug>=2)
logging.info("pose done")
if self.debug>=2:
extra['vis'] = vis
self.pose_last = pose
return (pose@self.get_tf_to_centered_mesh()).data.cpu().numpy().reshape(4,4)