-
Notifications
You must be signed in to change notification settings - Fork 20
/
vmap.py
525 lines (426 loc) · 21.2 KB
/
vmap.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
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
import random
import numpy as np
import torch
from time import perf_counter_ns
from tqdm import tqdm
import trainer
import open3d
import trimesh
import scipy
from bidict import bidict
import copy
import os
import utils
class performance_measure:
def __init__(self, name) -> None:
self.name = name
def __enter__(self):
self.start_time = perf_counter_ns()
def __exit__(self, type, value, tb):
self.end_time = perf_counter_ns()
self.exec_time = self.end_time - self.start_time
print(f"{self.name} excution time: {(self.exec_time)/1000000:.2f} ms")
def origin_dirs_W(T_WC, dirs_C):
assert T_WC.shape[0] == dirs_C.shape[0]
assert T_WC.shape[1:] == (4, 4)
assert dirs_C.shape[2] == 3
dirs_W = (T_WC[:, None, :3, :3] @ dirs_C[..., None]).squeeze()
origins = T_WC[:, :3, -1]
return origins, dirs_W
# @torch.jit.script
def stratified_bins(min_depth, max_depth, n_bins, n_rays, type=torch.float32, device = "cuda:0"):
# type: (Tensor, Tensor, int, int) -> Tensor
bin_limits_scale = torch.linspace(0, 1, n_bins+1, dtype=type, device=device)
if not torch.is_tensor(min_depth):
min_depth = torch.ones(n_rays, dtype=type, device=device) * min_depth
if not torch.is_tensor(max_depth):
max_depth = torch.ones(n_rays, dtype=type, device=device) * max_depth
depth_range = max_depth - min_depth
lower_limits_scale = depth_range[..., None] * bin_limits_scale + min_depth[..., None]
lower_limits_scale = lower_limits_scale[:, :-1]
assert lower_limits_scale.shape == (n_rays, n_bins)
bin_length_scale = depth_range / n_bins
increments_scale = torch.rand(
n_rays, n_bins, device=device,
dtype=torch.float32) * bin_length_scale[..., None]
z_vals_scale = lower_limits_scale + increments_scale
assert z_vals_scale.shape == (n_rays, n_bins)
return z_vals_scale
# @torch.jit.script
def normal_bins_sampling(depth, n_bins, n_rays, delta, device = "cuda:0"):
# type: (Tensor, int, int, float) -> Tensor
# device = "cpu"
# bins = torch.normal(0.0, delta / 3., size=[n_rays, n_bins], devi
# self.keyframes_batch = torch.empty(self.n_keyframes,ce=device).sort().values
bins = torch.empty(n_rays, n_bins, dtype=torch.float32, device=device).normal_(mean=0.,std=delta / 3.).sort().values
bins = torch.clip(bins, -delta, delta)
z_vals = depth[:, None] + bins
assert z_vals.shape == (n_rays, n_bins)
return z_vals
class sceneObject:
"""
object instance mapping,
updating keyframes, get training samples, optimizing MLP map
"""
def __init__(self, cfg, obj_id, rgb:torch.tensor, depth:torch.tensor, mask:torch.tensor, bbox_2d:torch.tensor, t_wc:torch.tensor, live_frame_id) -> None:
self.do_bg = cfg.do_bg
self.obj_id = obj_id
self.data_device = cfg.data_device
self.training_device = cfg.training_device
assert rgb.shape[:2] == depth.shape
assert rgb.shape[:2] == mask.shape
assert bbox_2d.shape == (4,)
assert t_wc.shape == (4, 4,)
if self.do_bg and self.obj_id == 0: # do seperate bg
self.obj_scale = cfg.bg_scale
self.hidden_feature_size = cfg.hidden_feature_size_bg
self.n_bins_cam2surface = cfg.n_bins_cam2surface_bg
self.keyframe_step = cfg.keyframe_step_bg
else:
self.obj_scale = cfg.obj_scale
self.hidden_feature_size = cfg.hidden_feature_size
self.n_bins_cam2surface = cfg.n_bins_cam2surface
self.keyframe_step = cfg.keyframe_step
self.frames_width = rgb.shape[0]
self.frames_height = rgb.shape[1]
self.min_bound = cfg.min_depth
self.max_bound = cfg.max_depth
self.n_bins = cfg.n_bins
self.n_unidir_funcs = cfg.n_unidir_funcs
self.surface_eps = cfg.surface_eps
self.stop_eps = cfg.stop_eps
self.n_keyframes = 1 # Number of keyframes
self.kf_pointer = None
self.keyframe_buffer_size = cfg.keyframe_buffer_size
self.kf_id_dict = bidict({live_frame_id:0})
self.kf_buffer_full = False
self.frame_cnt = 0 # number of frames taken in
self.lastest_kf_queue = []
self.bbox = torch.empty( # obj bounding bounding box in the frame
self.keyframe_buffer_size,
4,
device=self.data_device) # [u low, u high, v low, v high]
self.bbox[0] = bbox_2d
# RGB + pixel state batch
self.rgb_idx = slice(0, 3)
self.state_idx = slice(3, 4)
self.rgbs_batch = torch.empty(self.keyframe_buffer_size,
self.frames_width,
self.frames_height,
4,
dtype=torch.uint8,
device=self.data_device)
# Pixel states:
self.other_obj = 0 # pixel doesn't belong to obj
self.this_obj = 1 # pixel belong to obj
self.unknown_obj = 2 # pixel state is unknown
# Initialize first frame rgb and pixel state
self.rgbs_batch[0, :, :, self.rgb_idx] = rgb
self.rgbs_batch[0, :, :, self.state_idx] = mask[..., None]
self.depth_batch = torch.empty(self.keyframe_buffer_size,
self.frames_width,
self.frames_height,
dtype=torch.float32,
device=self.data_device)
# Initialize first frame's depth
self.depth_batch[0] = depth
self.t_wc_batch = torch.empty(
self.keyframe_buffer_size, 4, 4,
dtype=torch.float32,
device=self.data_device) # world to camera transform
# Initialize first frame's world2cam transform
self.t_wc_batch[0] = t_wc
# neural field map
trainer_cfg = copy.deepcopy(cfg)
trainer_cfg.obj_id = self.obj_id
trainer_cfg.hidden_feature_size = self.hidden_feature_size
trainer_cfg.obj_scale = self.obj_scale
self.trainer = trainer.Trainer(trainer_cfg)
# 3D boundary
self.bbox3d = None
self.pc = []
# init obj local frame
# self.obj_center = self.init_obj_center(intrinsic, depth, mask, t_wc)
self.obj_center = torch.tensor(0.0) # shouldn't make any difference because of frequency embedding
def init_obj_center(self, intrinsic_open3d, depth, mask, t_wc):
obj_depth = depth.cpu().clone()
obj_depth[mask!=self.this_obj] = 0
T_CW = np.linalg.inv(t_wc.cpu().numpy())
pc_obj_init = open3d.geometry.PointCloud.create_from_depth_image(
depth=open3d.geometry.Image(np.asarray(obj_depth.permute(1,0).numpy(), order="C")),
intrinsic=intrinsic_open3d,
extrinsic=T_CW,
depth_trunc=self.max_bound,
depth_scale=1.0)
obj_center = torch.from_numpy(np.mean(pc_obj_init.points, axis=0)).float()
return obj_center
# @profile
def append_keyframe(self, rgb:torch.tensor, depth:torch.tensor, mask:torch.tensor, bbox_2d:torch.tensor, t_wc:torch.tensor, frame_id:np.uint8=1):
assert rgb.shape[:2] == depth.shape
assert rgb.shape[:2] == mask.shape
assert bbox_2d.shape == (4,)
assert t_wc.shape == (4, 4,)
assert self.n_keyframes <= self.keyframe_buffer_size - 1
assert rgb.dtype == torch.uint8
assert mask.dtype == torch.uint8
assert depth.dtype == torch.float32
# every kf_step choose one kf
is_kf = (self.frame_cnt % self.keyframe_step == 0) or self.n_keyframes == 1
# print("---------------------")
# print("self.kf_id_dict ", self.kf_id_dict)
# print("live frame id ", frame_id)
# print("n_frames ", self.n_keyframes)
if self.n_keyframes == self.keyframe_buffer_size - 1: # kf buffer full, need to prune
self.kf_buffer_full = True
if self.kf_pointer is None:
self.kf_pointer = self.n_keyframes
self.rgbs_batch[self.kf_pointer, :, :, self.rgb_idx] = rgb
self.rgbs_batch[self.kf_pointer, :, :, self.state_idx] = mask[..., None]
self.depth_batch[self.kf_pointer, ...] = depth
self.t_wc_batch[self.kf_pointer, ...] = t_wc
self.bbox[self.kf_pointer, ...] = bbox_2d
self.kf_id_dict.inv[self.kf_pointer] = frame_id
if is_kf:
self.lastest_kf_queue.append(self.kf_pointer)
pruned_frame_id, pruned_kf_id = self.prune_keyframe()
self.kf_pointer = pruned_kf_id
print("pruned kf id ", self.kf_pointer)
else:
if not is_kf: # not kf, replace
self.rgbs_batch[self.n_keyframes-1, :, :, self.rgb_idx] = rgb
self.rgbs_batch[self.n_keyframes-1, :, :, self.state_idx] = mask[..., None]
self.depth_batch[self.n_keyframes-1, ...] = depth
self.t_wc_batch[self.n_keyframes-1, ...] = t_wc
self.bbox[self.n_keyframes-1, ...] = bbox_2d
self.kf_id_dict.inv[self.n_keyframes-1] = frame_id
else: # is kf, add new kf
self.kf_id_dict[frame_id] = self.n_keyframes
self.rgbs_batch[self.n_keyframes, :, :, self.rgb_idx] = rgb
self.rgbs_batch[self.n_keyframes, :, :, self.state_idx] = mask[..., None]
self.depth_batch[self.n_keyframes, ...] = depth
self.t_wc_batch[self.n_keyframes, ...] = t_wc
self.bbox[self.n_keyframes, ...] = bbox_2d
self.lastest_kf_queue.append(self.n_keyframes)
self.n_keyframes += 1
# print("self.kf_id_dic ", self.kf_id_dict)
self.frame_cnt += 1
if len(self.lastest_kf_queue) > 2: # keep latest two frames
self.lastest_kf_queue = self.lastest_kf_queue[-2:]
def prune_keyframe(self):
# simple strategy to prune, randomly choose
key, value = random.choice(list(self.kf_id_dict.items())[:-2]) # do not prune latest two frames
return key, value
def get_bound(self, intrinsic_open3d):
# get 3D boundary from posed depth img todo update sparse pc when append frame
pcs = open3d.geometry.PointCloud()
for kf_id in range(self.n_keyframes):
mask = self.rgbs_batch[kf_id, :, :, self.state_idx].squeeze() == self.this_obj
depth = self.depth_batch[kf_id].cpu().clone()
twc = self.t_wc_batch[kf_id].cpu().numpy()
depth[~mask] = 0
depth = depth.permute(1,0).numpy().astype(np.float32)
T_CW = np.linalg.inv(twc)
pc = open3d.geometry.PointCloud.create_from_depth_image(depth=open3d.geometry.Image(np.asarray(depth, order="C")), intrinsic=intrinsic_open3d, extrinsic=T_CW)
# self.pc += pc
pcs += pc
# # get minimal oriented 3d bbox
# try:
# bbox3d = open3d.geometry.OrientedBoundingBox.create_from_points(pcs.points)
# except RuntimeError:
# print("too few pcs obj ")
# return None
# trimesh has a better minimal bbox implementation than open3d
try:
transform, extents = trimesh.bounds.oriented_bounds(np.array(pcs.points)) # pc
transform = np.linalg.inv(transform)
except scipy.spatial._qhull.QhullError:
print("too few pcs obj ")
return None
for i in range(extents.shape[0]):
extents[i] = np.maximum(extents[i], 0.10) # at least rendering 10cm
bbox = utils.BoundingBox()
bbox.center = transform[:3, 3]
bbox.R = transform[:3, :3]
bbox.extent = extents
bbox3d = open3d.geometry.OrientedBoundingBox(bbox.center, bbox.R, bbox.extent)
min_extent = 0.05
bbox3d.extent = np.maximum(min_extent, bbox3d.extent)
bbox3d.color = (255,0,0)
self.bbox3d = utils.bbox_open3d2bbox(bbox_o3d=bbox3d)
# self.pc = []
print("obj ", self.obj_id)
print("bound ", bbox3d)
print("kf id dict ", self.kf_id_dict)
# open3d.visualization.draw_geometries([bbox3d, pcs])
return bbox3d
def get_training_samples(self, n_frames, n_samples, cached_rays_dir):
# Sample pixels
if self.n_keyframes > 2: # make sure latest 2 frames are sampled todo if kf pruned, this is not the latest frame
keyframe_ids = torch.randint(low=0,
high=self.n_keyframes,
size=(n_frames - 2,),
dtype=torch.long,
device=self.data_device)
# if self.kf_buffer_full:
# latest_frame_ids = list(self.kf_id_dict.values())[-2:]
latest_frame_ids = self.lastest_kf_queue[-2:]
keyframe_ids = torch.cat([keyframe_ids,
torch.tensor(latest_frame_ids, device=keyframe_ids.device)])
# print("latest_frame_ids", latest_frame_ids)
# else: # sample last 2 frames
# keyframe_ids = torch.cat([keyframe_ids,
# torch.tensor([self.n_keyframes-2, self.n_keyframes-1], device=keyframe_ids.device)])
else:
keyframe_ids = torch.randint(low=0,
high=self.n_keyframes,
size=(n_frames,),
dtype=torch.long,
device=self.data_device)
keyframe_ids = torch.unsqueeze(keyframe_ids, dim=-1)
idx_w = torch.rand(n_frames, n_samples, device=self.data_device)
idx_h = torch.rand(n_frames, n_samples, device=self.data_device)
# resizing idx_w and idx_h to be in the bbox range
idx_w = idx_w * (self.bbox[keyframe_ids, 1] - self.bbox[keyframe_ids, 0]) + self.bbox[keyframe_ids, 0]
idx_h = idx_h * (self.bbox[keyframe_ids, 3] - self.bbox[keyframe_ids, 2]) + self.bbox[keyframe_ids, 2]
idx_w = idx_w.long()
idx_h = idx_h.long()
sampled_rgbs = self.rgbs_batch[keyframe_ids, idx_w, idx_h]
sampled_depth = self.depth_batch[keyframe_ids, idx_w, idx_h]
# Get ray directions for sampled pixels
sampled_ray_dirs = cached_rays_dir[idx_w, idx_h]
# Get sampled keyframe poses
sampled_twc = self.t_wc_batch[keyframe_ids[:, 0], :, :]
origins, dirs_w = origin_dirs_W(sampled_twc, sampled_ray_dirs)
return self.sample_3d_points(sampled_rgbs, sampled_depth, origins, dirs_w)
def sample_3d_points(self, sampled_rgbs, sampled_depth, origins, dirs_w):
"""
3D sampling strategy
* For pixels with invalid depth:
- N+M from minimum bound to max (stratified)
* For pixels with valid depth:
# Pixel belongs to this object
- N from cam to surface (stratified)
- M around surface (stratified/normal)
# Pixel belongs that don't belong to this object
- N from cam to surface (stratified)
- M around surface (stratified)
# Pixel with unknown state
- Do nothing!
"""
n_bins_cam2surface = self.n_bins_cam2surface
n_bins = self.n_bins
eps = self.surface_eps
other_objs_max_eps = self.stop_eps #0.05 # todo 0.02
# print("max depth ", torch.max(sampled_depth))
sampled_z = torch.zeros(
sampled_rgbs.shape[0] * sampled_rgbs.shape[1],
n_bins_cam2surface + n_bins,
dtype=self.depth_batch.dtype,
device=self.data_device) # shape (N*n_rays, n_bins_cam2surface + n_bins)
invalid_depth_mask = (sampled_depth <= self.min_bound).view(-1)
# max_bound = self.max_bound
max_bound = torch.max(sampled_depth)
# sampling for points with invalid depth
invalid_depth_count = invalid_depth_mask.count_nonzero()
if invalid_depth_count:
sampled_z[invalid_depth_mask, :] = stratified_bins(
self.min_bound, max_bound,
n_bins_cam2surface + n_bins, invalid_depth_count,
device=self.data_device)
# sampling for valid depth rays
valid_depth_mask = ~invalid_depth_mask
valid_depth_count = valid_depth_mask.count_nonzero()
if valid_depth_count:
# Sample between min bound and depth for all pixels with valid depth
sampled_z[valid_depth_mask, :n_bins_cam2surface] = stratified_bins(
self.min_bound, sampled_depth.view(-1)[valid_depth_mask]-eps,
n_bins_cam2surface, valid_depth_count, device=self.data_device)
# sampling around depth for this object
obj_mask = (sampled_rgbs[..., -1] == self.this_obj).view(-1) & valid_depth_mask # todo obj_mask
assert sampled_z.shape[0] == obj_mask.shape[0]
obj_count = obj_mask.count_nonzero()
if obj_count:
sampling_method = "normal" # stratified or normal
if sampling_method == "stratified":
sampled_z[obj_mask, n_bins_cam2surface:] = stratified_bins(
sampled_depth.view(-1)[obj_mask] - eps, sampled_depth.view(-1)[obj_mask] + eps,
n_bins, obj_count, device=self.data_device)
elif sampling_method == "normal":
sampled_z[obj_mask, n_bins_cam2surface:] = normal_bins_sampling(
sampled_depth.view(-1)[obj_mask],
n_bins,
obj_count,
delta=eps,
device=self.data_device)
else:
raise (
f"sampling method not implemented {sampling_method}, \
stratified and normal sampling only currenty implemented."
)
# sampling around depth of other objects
other_obj_mask = (sampled_rgbs[..., -1] != self.this_obj).view(-1) & valid_depth_mask
other_objs_count = other_obj_mask.count_nonzero()
if other_objs_count:
sampled_z[other_obj_mask, n_bins_cam2surface:] = stratified_bins(
sampled_depth.view(-1)[other_obj_mask] - eps,
sampled_depth.view(-1)[other_obj_mask] + other_objs_max_eps,
n_bins, other_objs_count, device=self.data_device)
sampled_z = sampled_z.view(sampled_rgbs.shape[0],
sampled_rgbs.shape[1],
-1) # view as (n_rays, n_samples, 10)
input_pcs = origins[..., None, None, :] + (dirs_w[:, :, None, :] *
sampled_z[..., None])
input_pcs -= self.obj_center
obj_labels = sampled_rgbs[..., -1].view(-1)
return sampled_rgbs[..., :3], sampled_depth, valid_depth_mask, obj_labels, input_pcs, sampled_z
def save_checkpoints(self, path, epoch):
obj_id = self.obj_id
chechpoint_load_file = (path + "/obj_" + str(obj_id) + "_frame_" + str(epoch) + ".pth")
torch.save(
{
"epoch": epoch,
"FC_state_dict": self.trainer.fc_occ_map.state_dict(),
"PE_state_dict": self.trainer.pe.state_dict(),
"obj_id": self.obj_id,
"bbox": self.bbox3d,
"obj_scale": self.trainer.obj_scale
},
chechpoint_load_file,
)
# optimiser?
def load_checkpoints(self, ckpt_file):
checkpoint_load_file = (ckpt_file)
if not os.path.exists(checkpoint_load_file):
print("ckpt not exist ", checkpoint_load_file)
return
checkpoint = torch.load(checkpoint_load_file)
self.trainer.fc_occ_map.load_state_dict(checkpoint["FC_state_dict"])
self.trainer.pe.load_state_dict(checkpoint["PE_state_dict"])
self.obj_id = checkpoint["obj_id"]
self.bbox3d = checkpoint["bbox"]
self.trainer.obj_scale = checkpoint["obj_scale"]
self.trainer.fc_occ_map.to(self.training_device)
self.trainer.pe.to(self.training_device)
class cameraInfo:
def __init__(self, cfg) -> None:
self.device = cfg.data_device
self.width = cfg.W # Frame width
self.height = cfg.H # Frame height
self.fx = cfg.fx
self.fy = cfg.fy
self.cx = cfg.cx
self.cy = cfg.cy
self.rays_dir_cache = self.get_rays_dirs()
def get_rays_dirs(self, depth_type="z"):
idx_w = torch.arange(end=self.width, device=self.device)
idx_h = torch.arange(end=self.height, device=self.device)
dirs = torch.ones((self.width, self.height, 3), device=self.device)
dirs[:, :, 0] = ((idx_w - self.cx) / self.fx)[:, None]
dirs[:, :, 1] = ((idx_h - self.cy) / self.fy)
if depth_type == "euclidean":
raise Exception(
"Get camera rays directions with euclidean depth not yet implemented"
)
norm = torch.norm(dirs, dim=-1)
dirs = dirs * (1. / norm)[:, :, :, None]
return dirs