-
Notifications
You must be signed in to change notification settings - Fork 82
/
tool.py
439 lines (363 loc) · 15.6 KB
/
tool.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
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
import numpy as np
import os,sys,copy,glob,cv2,trimesh,time,shutil,pickle,gzip,logging,argparse,difflib
from sklearn.cluster import DBSCAN
logging.getLogger().setLevel(logging.FATAL)
code_dir = os.path.dirname(os.path.realpath(__file__))
sys.path.append(code_dir)
sys.path.append('{}/ss-pybullet'.format(code_dir))
import warnings
warnings.filterwarnings("ignore")
import open3d as o3d
from PIL import Image
from transformations import *
from pybullet_env.utils_pybullet import *
from pybullet_env.env import *
from pybullet_env.env_grasp import *
import pybullet as p
import pybullet_data
from Utils import *
from data_reader import *
import pybullet_tools.utils as PU
import matplotlib.pyplot as plt
from dexnet.grasping.gripper import RobotGripper
from autolab_core import YamlConfig
from dexnet.grasping.gripper import save_grasp_pose_mesh
from renderer import ModelRendererOffscreen
import mpl_toolkits.mplot3d.axes3d as p3
def compute_nunocs_label_worker(color_file):
color = np.array(Image.open(color_file))
depth = cv2.imread(color_file.replace('rgb','depth'),-1)/1e4
H,W = depth.shape[:2]
with open(color_file.replace('rgb.png','meta.pkl'),'rb') as ff:
meta = pickle.load(ff)
code_dir = os.path.dirname(os.path.realpath(__file__))
with open(f'{code_dir}/config.yml','r') as ff:
cfg = yaml.safe_load(ff)
mesh_pts = None
env_body_ids = meta['env_body_ids']
K = meta['K']
poses = meta['poses']
xyz_map = depth2xyzmap(depth,K)
seg = cv2.imread(color_file.replace('rgb','seg'),-1)
seg_ids = np.unique(seg)
nocs_image = np.zeros((H,W,3))
for seg_id in seg_ids:
if seg_id in env_body_ids:
continue
if np.sum(seg==seg_id)==0:
continue
if mesh_pts is None:
mesh_dir = meta['id_to_obj_file'][seg_id]
class_name = get_class_name(mesh_dir)
symmetry_tfs = get_symmetry_tfs(class_name,allow_reflection=False)
mesh_scale = meta['id_to_scales'][seg_id]
mesh = trimesh.load(mesh_dir)
mesh_pts = mesh.vertices.copy()*mesh_scale
max_xyz = mesh_pts.max(axis=0).reshape(1,3)
min_xyz = mesh_pts.min(axis=0).reshape(1,3)
center_xyz = (max_xyz+min_xyz)/2
valid_mask = (seg==seg_id) & (xyz_map[...,2]>=0.1)
tmp_xyz = xyz_map[valid_mask].reshape(-1,3)
ob_in_world = poses[seg_id].copy()
cam_in_world = meta['cam_in_world'].copy()
ob_in_cam = np.linalg.inv(cam_in_world)@ob_in_world
cam_in_ob = np.linalg.inv(ob_in_cam)
tmp_xyz = (cam_in_ob@to_homo(tmp_xyz).T).T[:,:3]
nunocs_scale = 1.
nocs_xyz = (tmp_xyz-center_xyz) / (max_xyz-min_xyz).reshape(1,3) #[-0.5,0.5]
nocs_xyz = np.clip(nocs_xyz,-0.5,0.5)
nocs_xyz /= nunocs_scale
nocs_image[valid_mask] = (nocs_xyz+0.5)*255
out_file = color_file.replace('rgb','nunocs')
nocs_image = np.clip(nocs_image,0,255)
nocs_image = nocs_image.round().astype(np.uint8)
Image.fromarray(nocs_image).save(out_file)
print(f"Write to {out_file}")
def compute_nunocs_label():
code_dir = os.path.dirname(os.path.realpath(__file__))
color_files = sorted(glob.glob(f'{code_dir}/dataset/{class_name}/**/*rgb.png',recursive=True))
print("#color_files={}".format(len(color_files)))
for color_file in color_files:
compute_nunocs_label_worker(color_file)
def fill_depth_normal_worker(reader,depth_file):
data = reader.read_data_by_colorfile(depth_file.replace('depth.png','rgb.png'),fetch=['xyz_map'])
valid_mask = data['xyz_map'][:,:,2]>=0.1
pts = data['xyz_map'][valid_mask].reshape(-1,3)
pcd = toOpen3dCloud(pts)
pcd.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.003, max_nn=30))
pcd = correct_pcd_normal_direction(pcd)
normals = np.asarray(pcd.normals).copy()
normal_map = np.zeros(data['xyz_map'].shape)
vs,us = np.where(valid_mask>0)
normal_map[vs,us] = normals
normal_map = np.round((normal_map+1)/2.0*255)
normal_map = np.clip(normal_map,0,255).astype(np.uint8)
out_file = depth_file.replace('depth','normal')
Image.fromarray(normal_map).save(out_file)
print(f"Write to {out_file}")
def fill_depth_normal():
reader = DataReader(cfg)
code_dir = os.path.dirname(os.path.realpath(__file__))
depth_files = sorted(glob.glob(f'{code_dir}/dataset/{class_name}/**/*depth.png',recursive=True))
print("#depth_files={}".format(len(depth_files)))
for depth_file in depth_files:
fill_depth_normal_worker(reader,depth_file)
def make_isolated_training_data_worker(cfg,color_file):
'''
Isolate objects in the scene
'''
print('color_file',color_file)
reader = DataReader(cfg)
data = reader.read_data_by_colorfile(color_file)
seg_ids = np.unique(data['seg'])
seg_ids = seg_ids[seg_ids>0]
for seg_id in seg_ids:
mask = (data['seg']==seg_id) & (data['depth']>=0.1)
if np.sum(mask)<100:
continue
cloud_xyz = data['xyz_map'][mask].reshape(-1,3)
cloud_nocs = data['nocs_map'][mask].reshape(-1,3)
cloud_rgb = data['rgb'][mask].reshape(-1,3)
cloud_normal = data['normal_map'][mask].reshape(-1,3)
out_data = {'cloud_xyz':cloud_xyz, 'cloud_nocs':cloud_nocs, 'cloud_rgb':cloud_rgb, 'cloud_normal':cloud_normal, 'color_file':color_file, 'seg_id':seg_id}
out_dir = color_file.replace('/train/','/train_isolated_nunocs/').replace('/test/','/test_isolated_nunocs/').replace('rgb.png','_seg{}.pkl'.format(seg_id))
os.system("mkdir -p {}".format(os.path.dirname(out_dir)))
with gzip.open(out_dir,'wb') as ff:
pickle.dump(out_data,ff)
def make_isolated_training_data():
code_dir = os.path.dirname(os.path.realpath(__file__))
color_files = sorted(glob.glob(f'{code_dir}/dataset/{class_name}/**/*rgb.png',recursive=True))
for color_file in color_files:
make_isolated_training_data_worker(cfg,color_file)
def make_crop_scene_dataset_worker(color_file,out_dir,reader,n_crop_per_side,downsample_size):
print(color_file)
index_str = re.findall(r'[0-9]{7}',color_file)[0]
data = reader.read_data_by_colorfile(color_file,fetch=['xyz_map','normal_map'])
valid_mask = data['depth']>=0.1
data = {
'cloud_xyz': data['xyz_map'][valid_mask].reshape(-1,3),
'cloud_rgb': data['rgb'][valid_mask].reshape(-1,3),
'cloud_normal': data['normal_map'][valid_mask].reshape(-1,3),
'cloud_seg': data['seg'][valid_mask].reshape(-1),
}
pcd = toOpen3dCloud(data['cloud_xyz'])
pcd = pcd.voxel_down_sample(voxel_size=downsample_size)
kdtree = cKDTree(np.asarray(pcd.points))
indices,dists = kdtree.query(data['cloud_xyz'])
max_xyz = data['cloud_xyz'].max(axis=0)
min_xyz = data['cloud_xyz'].min(axis=0)
crop_x_length = (max_xyz[0]-min_xyz[0])/n_crop_per_side
crop_y_length = (max_xyz[1]-min_xyz[1])/n_crop_per_side
print(f'crop_x_length',crop_x_length)
for x_crop in range(n_crop_per_side):
for y_crop in range(n_crop_per_side):
xmin = min_xyz[0]+x_crop*crop_x_length
xmax = xmin+crop_x_length+0.001
ymin = min_xyz[1]+y_crop*crop_y_length
ymax = ymin+crop_y_length+0.001
keep_mask = (data['cloud_xyz'][:,0]>=xmin) & (data['cloud_xyz'][:,0]<=xmax) & \
(data['cloud_xyz'][:,1]>=ymin) & (data['cloud_xyz'][:,1]<=ymax)
cropped_data = {}
for k in data.keys():
cropped_data[k] = data[k][keep_mask]
out_file = f'{out_dir}/{index_str}_x_crop_{x_crop}_y_crop_{y_crop}.pkl'
with gzip.open(out_file,'wb') as ff:
pickle.dump(cropped_data,ff)
if int(index_str)<5:
pcd = toOpen3dCloud(cropped_data['cloud_xyz'],cropped_data['cloud_rgb'],cropped_data['cloud_normal'])
o3d.io.write_point_cloud(out_file.replace('.pkl','.ply'),pcd)
def make_crop_scene_dataset():
'''For instance segmentation training, remove background e.g. bin
'''
reader = DataReader(cfg)
code_dir = os.path.dirname(os.path.realpath(__file__))
base_dir = f'{code_dir}/dataset/{class_name}'
for split in ['train','test']:
color_files = sorted(glob.glob(f'{base_dir}/{split}/*rgb.png',recursive=True))
print(f'color_files={len(color_files)}')
n_crop_per_side = 1
downsample_size = 0.0005
out_dir = f'{base_dir}/{split}_cloud_n_crop_per_side_{n_crop_per_side}_downsample_size_{round(downsample_size,5)}'
print(f'out_dir: {out_dir}')
os.system(f'rm -rf {out_dir} && mkdir -p {out_dir}')
for color_file in color_files:
make_crop_scene_dataset_worker(color_file,out_dir,reader,n_crop_per_side,downsample_size)
def compute_per_ob_visibility_worker(color_file,cfg):
out_file = color_file.replace('rgb.png','full_vis_mask.pkl')
if os.path.exists(out_file):
return
with open(color_file.replace('rgb.png','meta.pkl'),'rb') as ff:
meta = pickle.load(ff)
seg = cv2.imread(color_file.replace('rgb','seg'),-1)
seg_ids = np.unique(seg)
id_to_obj_file = meta['id_to_obj_file']
model_dir = None
for body_id in seg_ids:
if body_id in meta['env_body_ids']:
continue
if model_dir is None:
obj_file = meta['id_to_obj_file'][body_id]
mesh = trimesh.load(obj_file)
scale = meta['id_to_scales'][body_id]
mesh.vertices = mesh.vertices*scale.reshape(1,3)
model_dir = '/tmp/{}.obj'.format(os.path.basename(color_file).replace('.png',''))
mesh.export(model_dir)
break
K = np.array(meta['K']).reshape(3,3)
renderer = ModelRendererOffscreen([model_dir],K,H=cfg['H'],W=cfg['W'])
cam_in_world = meta['cam_in_world']
full_vis_mask = {}
for seg_id in seg_ids:
if seg_id in meta['env_body_ids']:
continue
ob_in_world = meta['poses'][seg_id]
ob_in_cam = np.linalg.inv(cam_in_world)@ob_in_world
color,depth = renderer.render([ob_in_cam])
vs,us = np.where(depth>=0.1)
full_vis_mask[seg_id] = np.concatenate((us.reshape(-1,1),vs.reshape(-1,1)),axis=1).astype(np.uint16)
with gzip.open(out_file,'wb') as ff:
pickle.dump(full_vis_mask,ff)
os.remove(model_dir)
def compute_per_ob_visibility():
code_dir = os.path.dirname(os.path.realpath(__file__))
color_files = sorted(glob.glob(f'{code_dir}/dataset/{class_name}/**/*rgb.png',recursive=True))
print('#color_files={}'.format(len(color_files)))
for color_file in color_files:
compute_per_ob_visibility_worker(color_file,cfg)
def make_dense_clutter_grasp_data_worker(color_file,cfg,gripper,grasps):
print('color_file',color_file)
with open(color_file.replace('rgb.png','meta.pkl'),'rb') as ff:
meta = pickle.load(ff)
K = np.array(meta['K']).reshape(3,3)
with gzip.open(color_file.replace('rgb.png','full_vis_mask.pkl'),'rb') as ff:
full_vis_mask = pickle.load(ff)
id_to_obj_file = meta['id_to_obj_file']
cam_in_world = meta['cam_in_world']
for body_id,obj_file in id_to_obj_file.items():
if body_id in meta['env_body_ids']:
continue
obj_file = id_to_obj_file[body_id]
scales = meta['id_to_scales'][body_id]
scales_tf = np.eye(4)
scales_tf[:3,:3] = np.diag(scales)
break
grasp_in_gripper = gripper.get_grasp_pose_in_gripper_base()
depth = cv2.imread(color_file.replace('rgb','depth'),-1)/1e4
xyz_map = depth2xyzmap(depth,K)
scene_pts = xyz_map[xyz_map[:,:,2]>=0.1].reshape(-1,3)
pcd = toOpen3dCloud(scene_pts)
pcd = pcd.voxel_down_sample(voxel_size=0.001)
scene_pts = np.asarray(pcd.points).copy()
seg = cv2.imread(color_file.replace('rgb','seg'),-1)
out_grasps = {}
check_finger_region = False
max_grasp_per_scene = 20
candidates = []
body_ids = np.unique(seg)
np.random.shuffle(body_ids)
n_rej = {
'dot': 0,
'gripper_region': 0,
'collision_with_scene': 0,
}
candidate_grasps = []
vis_ratio_dict = {}
for body_id in body_ids:
if body_id in meta['env_body_ids']:
continue
n_visible = (seg==body_id).sum()
n_full_vis = full_vis_mask[body_id].shape[0]
vis_ratio = n_visible/n_full_vis
vis_ratio_dict[body_id] = vis_ratio
ob_pts = xyz_map[seg==body_id].reshape(-1,3)
ob_pts = ob_pts[ob_pts[:,2]>=0.1]
pcd = toOpen3dCloud(ob_pts)
pcd = pcd.voxel_down_sample(voxel_size=0.001)
ob_pts = np.asarray(pcd.points)
if vis_ratio>=0.8:
grasp_ids = np.arange(len(grasps))
np.random.shuffle(grasp_ids)
for grasp_id in grasp_ids:
grasp = grasps[grasp_id]
grasp_pose = grasp.grasp_pose
grasp_pose = scales_tf@grasp_pose
grasp_pose = normalizeRotation(grasp_pose)
ob_in_world = meta['poses'][body_id]
ob_in_cam = np.linalg.inv(cam_in_world)@ob_in_world
grasp_in_cam = ob_in_cam@grasp_pose
approach_dir = (grasp_in_cam[:3,:3]@np.array([1,0,0]).reshape(3,1)).reshape(3)
dot = np.dot(approach_dir,np.array([0,0,1]))
if dot<0:
n_rej['dot'] += 1
continue
if check_finger_region:
ob_in_grasp = (np.linalg.inv(grasp_in_cam)@to_homo(ob_pts).T).T[:,:3]
valid_mask = (ob_in_grasp[:,0]>=gripper.finger_xmin) & (ob_in_grasp[:,0]<=gripper.finger_xmax) & (ob_in_grasp[:,1]>=gripper.finger_ymin) & (ob_in_grasp[:,1]<=gripper.finger_ymax) & (ob_in_grasp[:,2]>=gripper.finger_zmin) & (ob_in_grasp[:,2]<=gripper.finger_zmax)
if valid_mask.sum()==0:
n_rej['gripper_region'] += 1
continue
candidate_grasps.append([body_id,grasp_in_cam,grasp.perturbation_score])
continue
candidate_grasps = np.array(candidate_grasps)
print("candidate_grasps={}".format(len(candidate_grasps)))
ids = np.random.choice(len(candidate_grasps),size=min(max_grasp_per_scene,len(candidate_grasps)),replace=False)
candidate_grasps = candidate_grasps[ids]
out_grasps = {}
for (body_id,grasp_in_cam,grasp.perturbation_score) in candidate_grasps:
if body_id not in out_grasps:
out_grasps[body_id] = []
out_grasps[body_id].append([grasp_in_cam,grasp.perturbation_score])
msg = 'n_rej: '
for k,n in n_rej.items():
msg += f'{k}={n} '
print(msg)
out_file = color_file.replace('rgb.png','grasp.pkl')
with gzip.open(out_file,'wb') as ff:
pickle.dump(out_grasps,ff)
def make_dense_clutter_grasp_data():
code_dir = os.path.dirname(os.path.realpath(__file__))
color_files = sorted(glob.glob(f'{code_dir}/dataset/{class_name}/**/*rgb.png',recursive=True))
print("#color_files={}".format(len(color_files)))
names = cfg['dataset'][class_name]['train']
grasps_dict = {}
for name in names:
code_dir = os.path.dirname(os.path.realpath(__file__))
with gzip.open(f"{code_dir}/data/object_models/{name.replace('.obj','_grasp_balanced_score.pkl')}",'rb') as ff:
grasps = pickle.load(ff)
grasps_dict[name] = grasps
grasps_all = []
for i,color_file in enumerate(color_files):
if i%max(len(color_files)//10,1)==0:
print(f"Preparing grasps for each scene {i}/{len(color_files)}")
seg = cv2.imread(color_file.replace('rgb','seg'),-1)
seg_ids = np.unique(seg)
with open(color_file.replace('rgb.png','meta.pkl'),'rb') as ff:
meta = pickle.load(ff)
for body_id in seg_ids:
if body_id in meta['env_body_ids']:
continue
obj_file = meta['id_to_obj_file'][body_id]
ob_name = os.path.basename(obj_file)
break
grasps = np.array(grasps_dict[ob_name])
grasps = np.random.choice(grasps,size=min(100,len(grasps)),replace=False)
grasps_all.append(grasps)
del grasps_dict
for i in range(len(color_files)):
make_dense_clutter_grasp_data_worker(color_files[i],cfg,gripper,grasps_all[i])
if __name__=="__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--class_name',type=str,default='nut')
args = parser.parse_args()
code_dir = os.path.dirname(os.path.realpath(__file__))
with open('{}/config.yml'.format(code_dir),'r') as ff:
cfg = yaml.safe_load(ff)
cfg_grasp = YamlConfig("{}/config_grasp.yml".format(code_dir))
class_name = args.class_name
gripper = RobotGripper.load(gripper_dir=cfg_grasp['gripper_dir'][class_name])
compute_nunocs_label()
fill_depth_normal()
compute_per_ob_visibility()
make_isolated_training_data()
make_crop_scene_dataset()
make_dense_clutter_grasp_data()