-
Notifications
You must be signed in to change notification settings - Fork 17
/
hdata.py
388 lines (315 loc) · 11.8 KB
/
hdata.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
import traceback
from abc import ABC, abstractmethod
from typing import Dict, List
import warnings
import numpy as np
import torch
import os
import inspect
from ..utils.builder import build_transform
from ..utils.logger import logger
from ..utils.misc import CONST
from ..utils.transform import flip_2d, flip_3d, aa_to_rotmat, rotmat_to_aa
from torch.utils.data._utils.collate import default_collate\
class HDataset(ABC):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.data_mode = cfg.DATA_MODE
self.data_root = cfg.DATA_ROOT
self.data_split = cfg.DATA_SPLIT
self.use_cache = cfg.DATA_PRESET.USE_CACHE
self.use_full_image = bool(cfg.DATA_PRESET.USE_FULL_IMAGE)
self.bbox_expand_ratio = float(cfg.DATA_PRESET.BBOX_EXPAND_RATIO)
if self.use_full_image is True and self.bbox_expand_ratio != 1.0:
warnings.warn("When using full image, bbox expand ratio should be 1.0")
self.bbox_expand_ratio = 1.0
self.image_size = cfg.DATA_PRESET.IMAGE_SIZE # (W, H)
self.data_preset = cfg.DATA_PRESET
self.center_idx = int(cfg.DATA_PRESET.CENTER_IDX)
self.is_inference = bool(cfg.get("IS_INFERENCE", False))
self.root_assets = os.path.normpath("assets")
self.transform = build_transform(cfg=cfg.TRANSFORM, data_preset=self.data_preset)
logger.info(f"Initialized abstract class: HDataset")
def __len__(self):
return len(self)
@staticmethod
def load_dataset(self):
pass
@abstractmethod
def get_image(self, idx):
pass
@abstractmethod
def get_image_path(self, idx):
pass
@abstractmethod
def get_joints_3d(self, idx):
pass
@abstractmethod
def get_verts_3d(self, idx):
pass
@abstractmethod
def get_verts_uvd(self, idx):
pass
@abstractmethod
def get_joints_2d(self, idx):
pass
@abstractmethod
def get_joints_uvd(self, idx):
pass
@abstractmethod
def get_cam_intr(self, idx):
pass
@abstractmethod
def get_side(self, idx):
pass
@abstractmethod
def get_bbox_center_scale(self, idx):
pass
@abstractmethod
def get_sample_identifier(self, idx):
pass
@abstractmethod
def get_mano_pose(self, idx):
pass
@abstractmethod
def get_mano_shape(self, idx):
pass
def get_cam_center(self, idx):
intr = self.get_cam_intr(idx)
return np.array([intr[0, 2], intr[1, 2]])
# visible in raw image
def get_joints_2d_vis(self, joints_2d=None, img_size=None, **kwargs):
joints_vis = np.all((0 <= joints_2d) & (joints_2d < img_size), axis=1)
return joints_vis.astype(np.float32)
def getitem_2d(self, idx):
hand_side = self.get_side(idx)
bbox_center, bbox_scale = self.get_bbox_center_scale(idx)
bbox_scale = bbox_scale * self.bbox_expand_ratio # extend bbox sacle
joints_2d = self.get_joints_2d(idx)
image_path = self.get_image_path(idx)
image = self.get_image(idx)
raw_size = [image.shape[1], image.shape[0]] # (W, H)
joints_vis = self.get_joints_2d_vis(joints_2d=joints_2d, img_size=raw_size)
label = {
"idx": idx,
"bbox_center": bbox_center,
"bbox_scale": bbox_scale,
"joints_2d": joints_2d,
"joints_vis": joints_vis,
"image_path": image_path,
"raw_size": raw_size,
}
return image, label
def getitem_uvd(self, idx):
# Support FreiHAND, HO3D, DexYCB, YT3D, TMANO,
hand_side = self.get_side(idx)
bbox_center, bbox_scale = self.get_bbox_center_scale(idx)
bbox_scale = bbox_scale * self.bbox_expand_ratio # extend bbox sacle
verts_uvd = self.get_verts_uvd(idx)
joints_uvd = self.get_joints_uvd(idx)
joints_2d = self.get_joints_2d(idx)
image_path = self.get_image_path(idx)
image = self.get_image(idx)
raw_size = [image.shape[1], image.shape[0]] # (W, H)
joints_vis = self.get_joints_2d_vis(joints_2d=joints_2d, img_size=raw_size)
label = {
"idx": idx,
"bbox_center": bbox_center,
"bbox_scale": bbox_scale,
"joints_2d": joints_2d,
"verts_uvd": verts_uvd,
"joints_uvd": joints_uvd,
"joints_vis": joints_vis,
"image_path": image_path,
"raw_size": raw_size,
}
return image, label
def getitem_3d(self, idx):
# Support FreiHAND, HO3D, DexYCB
sample_id = self.get_sample_identifier(idx)
hand_side = self.get_side(idx)
bbox_center, bbox_scale = self.get_bbox_center_scale(idx)
bbox_scale = bbox_scale * self.bbox_expand_ratio # extend bbox sacle
cam_intr = self.get_cam_intr(idx)
cam_center = self.get_cam_center(idx)
joints_3d = self.get_joints_3d(idx)
verts_3d = self.get_verts_3d(idx)
joints_2d = self.get_joints_2d(idx)
image_path = self.get_image_path(idx)
mano_pose = self.get_mano_pose(idx)
mano_shape = self.get_mano_shape(idx)
image = self.get_image(idx)
raw_size = [image.shape[1], image.shape[0]] # (W, H)
joints_vis = self.get_joints_2d_vis(joints_2d=joints_2d, img_size=raw_size)
label = {
"idx": idx,
"sample_id": sample_id,
"cam_center": cam_center,
"bbox_center": bbox_center,
"bbox_scale": bbox_scale,
"cam_intr": cam_intr,
"joints_2d": joints_2d,
"joints_3d": joints_3d,
"verts_3d": verts_3d,
"joints_vis": joints_vis,
"mano_pose": mano_pose,
"mano_shape": mano_shape,
"image_path": image_path,
"raw_size": raw_size,
"hand_side": hand_side,
}
return image, label
def __getitem__(self, idx):
if self.data_mode not in ["2D", "UVD", "3D"]:
raise NotImplementedError(f"Unknown data mode: {self.data_mode}")
if self.data_mode == "2D":
image, label = self.getitem_2d(idx)
elif self.data_mode == "UVD":
image, label = self.getitem_uvd(idx)
elif self.data_mode == "3D":
image, label = self.getitem_3d(idx)
results = self.transform(image, label) # @NOTE data augmentation
results.update(label)
return results
class HODataset(HDataset, ABC):
def __init__(self, cfg):
super(HDataset, self).__init__(cfg)
logger.info(f"Initialized abstract class: HODataset")
@abstractmethod
def get_obj_id(self, idx):
pass
@abstractmethod
def get_obj_faces(self, idx):
pass
@abstractmethod
def get_obj_transf(self, idx):
pass
@abstractmethod
def get_obj_normals_3d(self, idx):
pass
@abstractmethod
def get_obj_verts_3d(self, idx):
pass
@abstractmethod
def get_obj_verts_can(self, idx):
pass
@abstractmethod
def get_obj_normals_can(self, idx):
pass
def get_processed_contact_info(self, idx):
return {}
def getitem_3d_hand_obj(self, idx):
image, label = self.getitem_3d(idx)
label["obj_verts_can"] = self.get_obj_verts_can(idx)
label["obj_id"] = self.get_obj_id(idx)
label["obj_verts_3d"] = self.get_obj_verts_3d(idx)
label["obj_normals_3d"] = self.get_obj_normals_3d(idx)
label["obj_transf"] = self.get_obj_transf(idx)
label["obj_faces"] = self.get_obj_faces(idx)
label["obj_verts_color"] = self.get_obj_verts_color(idx)
return image, label
def __getitem__(self, idx):
if self.data_mode not in ["2d", "uvd", "3d", "3d_hand_obj", "3d_hand_obj_contact"]:
raise NotImplementedError(f"Unknown data mode: {self.data_mode}")
if self.data_mode == "2d":
image, label = self.getitem_2d(idx)
elif self.data_mode == "uvd":
image, label = self.getitem_uvd(idx)
elif self.data_mode == "3d":
image, label = self.getitem_3d(idx)
elif self.data_mode == "3d_hand_obj":
image, label = self.getitem_3d_hand_obj(idx)
elif self.data_mode == "3d_hand_obj_contact":
image, label = self.getitem_3d_hand_obj(idx)
contact_label = self.get_processed_contact_info(idx)
label.update(contact_label)
results = self.transform(image, label) # @NOTE data augmentation
results.update(label)
return results
def ho_data_collate(batch: List[Dict]):
"""
Collate function, duplicating the items in extend_queries along the
first dimension so that they all have the same length.
Typically applies to faces and vertices, which have different sizes
depending on the object.
"""
extend_queries = {
# before aug
"obj_verts_can",
"obj_normals_can",
"obj_verts_3d",
"obj_normals_3d",
"obj_faces",
# after aug
"target_obj_verts_3d",
"target_obj_normals_3d",
# contact query
"vertex_contact",
"contact_region_id",
"anchor_id",
"anchor_dist",
"anchor_elasti",
"anchor_padding_mask"
}
pop_queries = []
for poppable_query in extend_queries:
if poppable_query in batch[0]:
pop_queries.append(poppable_query)
# Remove fields that don't have matching sizes
for pop_query in pop_queries:
padding_query_field = match_collate_queries(pop_query)
max_size = max([sample[pop_query].shape[0] for sample in batch])
for sample in batch:
pop_value = sample[pop_query]
orig_len = pop_value.shape[0]
# Repeat vertices so all have the same number
pop_value = np.concatenate([pop_value] * int(max_size / pop_value.shape[0] + 1))[:max_size]
sample[pop_query] = pop_value
if padding_query_field not in sample:
# generate a new field, contains padding mask
# note that only the beginning pop_value.shape[0] points are in effect
# so the mask will be a vector of length max_size, with origin_len ones in the beginning
padding_mask = np.zeros(max_size, dtype=int)
padding_mask[:orig_len] = 1
sample[padding_query_field] = padding_mask
# store the mask filtering the points
batch = default_collate(batch) # this function np -> torch
return batch
def match_collate_queries(query_spin):
object_vertex_queries = [
# before aug
"obj_verts_can",
"obj_normals_can",
"obj_verts_3d",
"obj_normals_3d",
# after aug
"target_obj_verts_3d",
"target_obj_normals_3d",
# contact query
"vertex_contact",
"contact_region_id",
"anchor_id",
"anchor_dist",
"anchor_elasti",
"anchor_padding_mask"
]
object_face_quries = [
"obj_faces",
]
if query_spin in object_vertex_queries:
return "obj_verts_padding_mask"
elif query_spin in object_face_quries:
return "obj_faces_padding_mask"
def auto_delegate_get(incls): # incls must be a HDataset class
def decorator(cls):
for name, member in inspect.getmembers(incls, predicate=inspect.isfunction):
if name.startswith("get_"):
def wrapped_get(self, idx, _name=name):
if self.inlier_indices is not None:
return getattr(self.base_dataset, _name)(self.inlier_indices[idx])
else:
return getattr(self.base_dataset, _name)(idx)
setattr(cls, name, wrapped_get)
return cls
return decorator