forked from electech6/NeRF-Based-SLAM-Incredible-Insights
-
Notifications
You must be signed in to change notification settings - Fork 0
/
coslam.py
800 lines (635 loc) · 41.8 KB
/
coslam.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
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
import os
#os.environ['TCNN_CUDA_ARCHITECTURES'] = '86'
# Package imports
import torch
import torch.optim as optim
import numpy as np
import random
import torch.nn.functional as F
import argparse
import shutil
import json
import cv2
from torch.utils.data import DataLoader
from tqdm import tqdm
# Local imports
import config
from model.scene_rep import JointEncoding
from model.keyframe import KeyFrameDatabase
from datasets.dataset import get_dataset
from utils import coordinates, extract_mesh, colormap_image
from tools.eval_ate import pose_evaluation
from optimization.utils import at_to_transform_matrix, qt_to_transform_matrix, matrix_to_axis_angle, matrix_to_quaternion
class CoSLAM():
def __init__(self, config):
self.config = config
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.dataset = get_dataset(config)
self.create_bounds() # 建图的边界
self.create_pose_data() # 创建存储 估计的位姿 和 数据集中的位姿gt 到用的字典
self.get_pose_representation() # 查看当前数据集是用轴角还是四元数表示的,tum数据集是轴角
self.keyframeDatabase = self.create_kf_database(config) # tum/fr1_desk: 每5帧取为一个关键帧
# ! -------------------- 1. Scene representation --------------------
self.model = JointEncoding(config, self.bounding_box).to(self.device) # 得到encoding/decoding网络,用于获得深度和颜色信息
def seed_everything(self, seed):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
def get_pose_representation(self):
'''
Get the pose representation axis-angle or quaternion
'''
if self.config['training']['rot_rep'] == 'axis_angle':
self.matrix_to_tensor = matrix_to_axis_angle
self.matrix_from_tensor = at_to_transform_matrix
print('Using axis-angle as rotation representation, identity init would cause inf')
elif self.config['training']['rot_rep'] == "quat":
print("Using quaternion as rotation representation")
self.matrix_to_tensor = matrix_to_quaternion
self.matrix_from_tensor = qt_to_transform_matrix
else:
raise NotImplementedError
def create_pose_data(self):
'''
Create the pose data
'''
self.est_c2w_data = {}
self.est_c2w_data_rel = {}
self.load_gt_pose()
def create_bounds(self):
'''
Get the pre-defined bounds for the scene
'''
self.bounding_box = torch.from_numpy(np.array(self.config['mapping']['bound'])).to(self.device)
self.marching_cube_bound = torch.from_numpy(np.array(self.config['mapping']['marching_cubes_bound'])).to(self.device)
def create_kf_database(self, config):
'''
Create the keyframe database
'''
num_kf = int(self.dataset.num_frames // self.config['mapping']['keyframe_every'] + 1)
print('#kf:', num_kf)
print('#Pixels to save:', self.dataset.num_rays_to_save)
return KeyFrameDatabase(config,
self.dataset.H,
self.dataset.W,
num_kf,
self.dataset.num_rays_to_save,
self.device)
def load_gt_pose(self):
'''
Load the ground truth pose
'''
self.pose_gt = {}
for i, pose in enumerate(self.dataset.poses):
self.pose_gt[i] = pose
def save_state_dict(self, save_path):
torch.save(self.model.state_dict(), save_path)
def load(self, load_path):
self.model.load_state_dict(torch.load(load_path))
def save_ckpt(self, save_path):
'''
Save the model parameters and the estimated pose
'''
save_dict = {'pose': self.est_c2w_data,
'pose_rel': self.est_c2w_data_rel,
'model': self.model.state_dict()}
torch.save(save_dict, save_path)
print('Save the checkpoint')
def load_ckpt(self, load_path):
'''
Load the model parameters and the estimated pose
'''
dict = torch.load(load_path)
self.model.load_state_dict(dict['model'])
self.est_c2w_data = dict['pose']
self.est_c2w_data_rel = dict['pose_rel']
def select_samples(self, H, W, samples):
'''
randomly select samples from the image
'''
#indice = torch.randint(H*W, (samples,))
indice = random.sample(range(H * W), int(samples)) # 从一个范围内的整数(0 到 H * W - 1)中随机选择samples个整数索引
indice = torch.tensor(indice)
return indice
def get_loss_from_ret(self, ret, rgb=True, sdf=True, depth=True, fs=True, smooth=False):
'''
Get the training loss
'''
loss = 0
if rgb:
loss += self.config['training']['rgb_weight'] * ret['rgb_loss']
if depth:
loss += self.config['training']['depth_weight'] * ret['depth_loss']
if sdf:
loss += self.config['training']['sdf_weight'] * ret["sdf_loss"]
if fs:
loss += self.config['training']['fs_weight'] * ret["fs_loss"]
if smooth and self.config['training']['smooth_weight']>0:
loss += self.config['training']['smooth_weight'] * self.smoothness(self.config['training']['smooth_pts'],
self.config['training']['smooth_vox'],
margin=self.config['training']['smooth_margin'])
return loss
def first_frame_mapping(self, batch, n_iters=100):
'''
First frame mapping
Params:
batch['c2w']: [1, 4, 4]
batch['rgb']: [1, H, W, 3]
batch['depth']: [1, H, W, 1]
batch['direction']: [1, H, W, 3]
Returns:
ret: dict
loss: float
'''
# ********************* 读取第0帧的相机位姿 *********************
print('First frame mapping...')
c2w = batch['c2w'][0].to(self.device) # [4,4]
self.est_c2w_data[0] = c2w # 第0帧的观测位姿 直接作为 位姿估计
self.est_c2w_data_rel[0] = c2w
self.model.train() # 将模型设置为训练模式
# Training n_iters=100
for i in range(n_iters):
# ********************* 获得第0帧每个像素的颜色,深度,方向 *********************
self.map_optimizer.zero_grad() # 将之前的梯度信息清零
indice = self.select_samples(self.dataset.H, self.dataset.W, self.config['mapping']['sample']) # 从一个范围内的整数(0 到 H * W - 1)中随机选择samples=2048个样本像素
indice_h, indice_w = indice % (self.dataset.H), indice // (self.dataset.H) # 得到每个采样得到的像素的h,w值 [2048]
rays_d_cam = batch['direction'].squeeze(0)[indice_h, indice_w, :].to(self.device) # 得到每个样本像素的方向,作为目标射线方向 [2048,3]
target_s = batch['rgb'].squeeze(0)[indice_h, indice_w, :].to(self.device) # 得到每个样本像素的颜色,作为目标颜色 [2048.3]
target_d = batch['depth'].squeeze(0)[indice_h, indice_w].to(self.device).unsqueeze(-1) # 得到每个样本像素的深度,作为目标深度 [2048,1]
rays_o = c2w[None, :3, -1].repeat(self.config['mapping']['sample'], 1) # 世界坐标系下的射线原点,即变换矩阵的t,即相机位置 [2048,3]
# rays_d_cam[..., None, :] 相机坐标系中的射线方向: [2048,1,3]
# c2w[:3, :3] : 旋转矩阵 [3,3]
# sum(x,-1) 在最后一个维度上进行求和
rays_d = torch.sum(rays_d_cam[..., None, :] * c2w[:3, :3], -1) # 世界坐标系下的射线方向 [2048,3]
# Forward
# ********************* 前向传播: 得到rgb图,深度图,rgb损失,深度损失,sdf损失,fs损失 *********************
ret = self.model.forward(rays_o, rays_d, target_s, target_d) # 前向传播
loss = self.get_loss_from_ret(ret) # 将所有损失求和
# ********************* 反响传播: 优化encoder/decoder网络的参数 *********************
loss.backward() # 反响传播
self.map_optimizer.step() # 使用adam优化encoder和decoder网络的参数
# ********************* 将当前帧加入关键帧 *********************
# First frame will always be a keyframe
self.keyframeDatabase.add_keyframe(batch, filter_depth=self.config['mapping']['filter_depth'])
if self.config['mapping']['first_mesh']:
self.save_mesh(0)
print('First frame mapping done')
return ret, loss # 返回: rgb图,深度图,rgb损失,深度损失,sdf损失,fs损失
def current_frame_mapping(self, batch, cur_frame_id):
'''
Current frame mapping
Params:
batch['c2w']: [1, 4, 4]
batch['rgb']: [1, H, W, 3]
batch['depth']: [1, H, W, 1]
batch['direction']: [1, H, W, 3]
Returns:
ret: dict
loss: float
'''
if self.config['mapping']['cur_frame_iters'] <= 0:
return
print('Current frame mapping...')
# ********************* 读取当前帧的位姿估计 *********************
c2w = self.est_c2w_data[cur_frame_id].to(self.device)
self.model.train() # 将模型设置为训练模式
# Training
for i in range(self.config['mapping']['cur_frame_iters']):
# ********************* 从数据集获得当前帧每个像素的颜色,深度,方向观测值 *********************
self.cur_map_optimizer.zero_grad() # 将之前的梯度信息清零
indice = self.select_samples(self.dataset.H, self.dataset.W, self.config['mapping']['sample'])
indice_h, indice_w = indice % (self.dataset.H), indice // (self.dataset.H)
rays_d_cam = batch['direction'].squeeze(0)[indice_h, indice_w, :].to(self.device)
target_s = batch['rgb'].squeeze(0)[indice_h, indice_w, :].to(self.device)
target_d = batch['depth'].squeeze(0)[indice_h, indice_w].to(self.device).unsqueeze(-1)
# ********************* 根据位姿估计计算射线的原点和方向 *********************
rays_o = c2w[None, :3, -1].repeat(self.config['mapping']['sample'], 1)
rays_d = torch.sum(rays_d_cam[..., None, :] * c2w[:3, :3], -1)
# ********************* 前向传播: 得到rgb图,深度图,rgb损失,深度损失,sdf损失,fs损失 *********************
ret = self.model.forward(rays_o, rays_d, target_s, target_d)
loss = self.get_loss_from_ret(ret) # 将所有损失求和
# ********************* 反响传播: 优化encoder/decoder网络的参数 *********************
loss.backward()
self.cur_map_optimizer.step()
return ret, loss # 返回: rgb图,深度图,rgb损失,深度损失,sdf损失,fs损失
def smoothness(self, sample_points=256, voxel_size=0.1, margin=0.05, color=False):
'''
Smoothness loss of feature grid
'''
volume = self.bounding_box[:, 1] - self.bounding_box[:, 0]
grid_size = (sample_points-1) * voxel_size
offset_max = self.bounding_box[:, 1]-self.bounding_box[:, 0] - grid_size - 2 * margin
offset = torch.rand(3).to(offset_max) * offset_max + margin
coords = coordinates(sample_points - 1, 'cpu', flatten=False).float().to(volume)
pts = (coords + torch.rand((1,1,1,3)).to(volume)) * voxel_size + self.bounding_box[:, 0] + offset
if self.config['grid']['tcnn_encoding']:
pts_tcnn = (pts - self.bounding_box[:, 0]) / (self.bounding_box[:, 1] - self.bounding_box[:, 0])
sdf = self.model.query_sdf(pts_tcnn, embed=True)
tv_x = torch.pow(sdf[1:,...]-sdf[:-1,...], 2).sum()
tv_y = torch.pow(sdf[:,1:,...]-sdf[:,:-1,...], 2).sum()
tv_z = torch.pow(sdf[:,:,1:,...]-sdf[:,:,:-1,...], 2).sum()
loss = (tv_x + tv_y + tv_z)/ (sample_points**3)
return loss
def get_pose_param_optim(self, poses, mapping=True):
task = 'mapping' if mapping else 'tracking'
cur_trans = torch.nn.parameter.Parameter(poses[:, :3, 3]) # 提取变换矩阵的t
cur_rot = torch.nn.parameter.Parameter(self.matrix_to_tensor(poses[:, :3, :3])) # 提取变换矩阵的R
pose_optimizer = torch.optim.Adam([{"params": cur_rot, "lr": self.config[task]['lr_rot']},
{"params": cur_trans, "lr": self.config[task]['lr_trans']}])
return cur_rot, cur_trans, pose_optimizer
def global_BA(self, batch, cur_frame_id):
'''
Global bundle adjustment that includes all the keyframes and the current frame
Params:
batch['c2w']: ground truth camera pose [1, 4, 4]
batch['rgb']: rgb image [1, H, W, 3]
batch['depth']: depth image [1, H, W, 1]
batch['direction']: view direction [1, H, W, 3]
cur_frame_id: current frame id
'''
pose_optimizer = None
# all the KF poses: 0, 5, 10, ...
poses = torch.stack([self.est_c2w_data[i] for i in range(0, cur_frame_id, self.config['mapping']['keyframe_every'])])
# frame ids for all KFs, used for update poses after optimization
frame_ids_all = torch.tensor(list(range(0, cur_frame_id, self.config['mapping']['keyframe_every'])))
if len(self.keyframeDatabase.frame_ids) < 2:
poses_fixed = torch.nn.parameter.Parameter(poses).to(self.device)
current_pose = self.est_c2w_data[cur_frame_id][None,...]
poses_all = torch.cat([poses_fixed, current_pose], dim=0)
else:
poses_fixed = torch.nn.parameter.Parameter(poses[:1]).to(self.device)
current_pose = self.est_c2w_data[cur_frame_id][None,...]
if self.config['mapping']['optim_cur']:
cur_rot, cur_trans, pose_optimizer, = self.get_pose_param_optim(torch.cat([poses[1:], current_pose]))
pose_optim = self.matrix_from_tensor(cur_rot, cur_trans).to(self.device)
poses_all = torch.cat([poses_fixed, pose_optim], dim=0)
else:
cur_rot, cur_trans, pose_optimizer, = self.get_pose_param_optim(poses[1:])
pose_optim = self.matrix_from_tensor(cur_rot, cur_trans).to(self.device)
poses_all = torch.cat([poses_fixed, pose_optim, current_pose], dim=0)
# Set up optimizer
self.map_optimizer.zero_grad()
if pose_optimizer is not None:
pose_optimizer.zero_grad()
current_rays = torch.cat([batch['direction'], batch['rgb'], batch['depth'][..., None]], dim=-1)
current_rays = current_rays.reshape(-1, current_rays.shape[-1])
for i in range(self.config['mapping']['iters']):
# Sample rays with real frame ids
# rays [bs, 7]
# frame_ids [bs]
rays, ids = self.keyframeDatabase.sample_global_rays(self.config['mapping']['sample'])
#TODO: Checkpoint...
idx_cur = random.sample(range(0, self.dataset.H * self.dataset.W),max(self.config['mapping']['sample'] // len(self.keyframeDatabase.frame_ids), self.config['mapping']['min_pixels_cur']))
current_rays_batch = current_rays[idx_cur, :]
rays = torch.cat([rays, current_rays_batch], dim=0) # N, 7
ids_all = torch.cat([ids//self.config['mapping']['keyframe_every'], -torch.ones((len(idx_cur)))]).to(torch.int64)
rays_d_cam = rays[..., :3].to(self.device)
target_s = rays[..., 3:6].to(self.device)
target_d = rays[..., 6:7].to(self.device)
# [N, Bs, 1, 3] * [N, 1, 3, 3] = (N, Bs, 3)
rays_d = torch.sum(rays_d_cam[..., None, None, :] * poses_all[ids_all, None, :3, :3], -1)
rays_o = poses_all[ids_all, None, :3, -1].repeat(1, rays_d.shape[1], 1).reshape(-1, 3)
rays_d = rays_d.reshape(-1, 3)
ret = self.model.forward(rays_o, rays_d, target_s, target_d)
loss = self.get_loss_from_ret(ret, smooth=True)
loss.backward(retain_graph=True)
if (i + 1) % cfg["mapping"]["map_accum_step"] == 0:
if (i + 1) > cfg["mapping"]["map_wait_step"]:
self.map_optimizer.step()
else:
print('Wait update')
self.map_optimizer.zero_grad()
if pose_optimizer is not None and (i + 1) % cfg["mapping"]["pose_accum_step"] == 0:
pose_optimizer.step()
# get SE3 poses to do forward pass
pose_optim = self.matrix_from_tensor(cur_rot, cur_trans)
pose_optim = pose_optim.to(self.device)
# So current pose is always unchanged
if self.config['mapping']['optim_cur']:
poses_all = torch.cat([poses_fixed, pose_optim], dim=0)
else:
current_pose = self.est_c2w_data[cur_frame_id][None,...]
# SE3 poses
poses_all = torch.cat([poses_fixed, pose_optim, current_pose], dim=0)
# zero_grad here
pose_optimizer.zero_grad()
if pose_optimizer is not None and len(frame_ids_all) > 1:
for i in range(len(frame_ids_all[1:])):
self.est_c2w_data[int(frame_ids_all[i+1].item())] = self.matrix_from_tensor(cur_rot[i:i+1], cur_trans[i:i+1]).detach().clone()[0]
if self.config['mapping']['optim_cur']:
print('Update current pose')
self.est_c2w_data[cur_frame_id] = self.matrix_from_tensor(cur_rot[-1:], cur_trans[-1:]).detach().clone()[0]
def predict_current_pose(self, frame_id, constant_speed=True):
'''
Predict current pose from previous pose using camera motion model
'''
if frame_id == 1 or (not constant_speed): # 第0帧已用于初始训练encoder和decoder网络
c2w_est_prev = self.est_c2w_data[frame_id-1].to(self.device) # 此时,读取第0帧的位姿
self.est_c2w_data[frame_id] = c2w_est_prev # 作为第1帧的位姿估计
else:
# ! -------------------- 2.2 Camera tracking: 初始化位姿估计 --------------------
# 对于后面的帧使用论文的公式10,来初始化当前帧的位姿估计
c2w_est_prev_prev = self.est_c2w_data[frame_id-2].to(self.device) # 第i-2帧的位姿
c2w_est_prev = self.est_c2w_data[frame_id-1].to(self.device) # 第i-1帧的位姿
delta = c2w_est_prev@c2w_est_prev_prev.float().inverse() # T1 * T2^-1
self.est_c2w_data[frame_id] = delta@c2w_est_prev # T1 * T2^-1 * T1
return self.est_c2w_data[frame_id]
def tracking_pc(self, batch, frame_id):
'''
Tracking camera pose of current frame using point cloud loss
(Not used in the paper, but might be useful for some cases)
'''
c2w_gt = batch['c2w'][0].to(self.device)
cur_c2w = self.predict_current_pose(frame_id, self.config['tracking']['const_speed'])
cur_trans = torch.nn.parameter.Parameter(cur_c2w[..., :3, 3].unsqueeze(0))
cur_rot = torch.nn.parameter.Parameter(self.matrix_to_tensor(cur_c2w[..., :3, :3]).unsqueeze(0))
pose_optimizer = torch.optim.Adam([{"params": cur_rot, "lr": self.config['tracking']['lr_rot']},
{"params": cur_trans, "lr": self.config['tracking']['lr_trans']}])
best_sdf_loss = None
iW = self.config['tracking']['ignore_edge_W']
iH = self.config['tracking']['ignore_edge_H']
thresh=0
if self.config['tracking']['iter_point'] > 0:
indice_pc = self.select_samples(self.dataset.H-iH*2, self.dataset.W-iW*2, self.config['tracking']['pc_samples'])
rays_d_cam = batch['direction'][:, iH:-iH, iW:-iW].reshape(-1, 3)[indice_pc].to(self.device)
target_s = batch['rgb'][:, iH:-iH, iW:-iW].reshape(-1, 3)[indice_pc].to(self.device)
target_d = batch['depth'][:, iH:-iH, iW:-iW].reshape(-1, 1)[indice_pc].to(self.device)
valid_depth_mask = ((target_d > 0.) * (target_d < 5.))[:,0]
rays_d_cam = rays_d_cam[valid_depth_mask]
target_s = target_s[valid_depth_mask]
target_d = target_d[valid_depth_mask]
for i in range(self.config['tracking']['iter_point']):
pose_optimizer.zero_grad()
c2w_est = self.matrix_from_tensor(cur_rot, cur_trans)
rays_o = c2w_est[...,:3, -1].repeat(len(rays_d_cam), 1)
rays_d = torch.sum(rays_d_cam[..., None, :] * c2w_est[:, :3, :3], -1)
pts = rays_o + target_d * rays_d
pts_flat = (pts - self.bounding_box[:, 0]) / (self.bounding_box[:, 1] - self.bounding_box[:, 0])
out = self.model.query_color_sdf(pts_flat)
sdf = out[:, -1]
rgb = torch.sigmoid(out[:,:3])
# TODO: Change this
loss = 5 * torch.mean(torch.square(rgb-target_s)) + 1000 * torch.mean(torch.square(sdf))
if best_sdf_loss is None:
best_sdf_loss = loss.cpu().item()
best_c2w_est = c2w_est.detach()
with torch.no_grad():
c2w_est = self.matrix_from_tensor(cur_rot, cur_trans)
if loss.cpu().item() < best_sdf_loss:
best_sdf_loss = loss.cpu().item()
best_c2w_est = c2w_est.detach()
thresh = 0
else:
thresh +=1
if thresh >self.config['tracking']['wait_iters']:
break
loss.backward()
pose_optimizer.step()
if self.config['tracking']['best']:
self.est_c2w_data[frame_id] = best_c2w_est.detach().clone()[0]
else:
self.est_c2w_data[frame_id] = c2w_est.detach().clone()[0]
if frame_id % self.config['mapping']['keyframe_every'] != 0:
# Not a keyframe, need relative pose
kf_id = frame_id // self.config['mapping']['keyframe_every']
kf_frame_id = kf_id * self.config['mapping']['keyframe_every']
c2w_key = self.est_c2w_data[kf_frame_id]
delta = self.est_c2w_data[frame_id] @ c2w_key.float().inverse()
self.est_c2w_data_rel[frame_id] = delta
print('Best loss: {}, Camera loss{}'.format(F.l1_loss(best_c2w_est.to(self.device)[0,:3], c2w_gt[:3]).cpu().item(), F.l1_loss(c2w_est[0,:3], c2w_gt[:3]).cpu().item()))
def tracking_render(self, batch, frame_id):
'''
Tracking camera pose using of the current frame
Params:
batch['c2w']: Ground truth camera pose [B, 4, 4]
batch['rgb']: RGB image [B, H, W, 3]
batch['depth']: Depth image [B, H, W, 1]
batch['direction']: Ray direction [B, H, W, 3]
frame_id: Current frame id (int)
'''
c2w_gt = batch['c2w'][0].to(self.device) # 从数据集得到当前帧的位姿真值 [4, 4]
# ********************* 初始化当前帧的位姿估计 *********************
# Initialize current pose
if self.config['tracking']['iter_point'] > 0: # 通过点云损失来跟踪当前帧的相机位姿时使用,本论文没用该方法
cur_c2w = self.est_c2w_data[frame_id]
else: # 使用论文的方法来跟踪相机位姿时使用
# ! -------------------- 2.2 Camera tracking: 初始化位姿估计 --------------------
cur_c2w = self.predict_current_pose(frame_id, self.config['tracking']['const_speed'])
# ********************* 一些训练变量 *********************
indice = None
best_sdf_loss = None
thresh=0
iW = self.config['tracking']['ignore_edge_W'] # 20
iH = self.config['tracking']['ignore_edge_H'] # 20
# ********************* 为位姿估计设置优化器 *********************
# 提取当前位姿的t,R。并给他们创建Adam优化器
# cur_rot: [1,3] 转为轴角表示
# cur_trans: [1,3] 位移
cur_rot, cur_trans, pose_optimizer = self.get_pose_param_optim(cur_c2w[None,...], mapping=False)
# ! -------------------- 2.2 Camera tracking: 优化位姿 --------------------
# >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
# ********************* 开始跟踪 *********************
# Start tracking
for i in range(self.config['tracking']['iter']): # iter = 10
pose_optimizer.zero_grad() # 将R和t的adam优化器梯度设置为0
# ********************* 将位姿估计:cur_rot, cur_trans转为变换矩阵 *********************
c2w_est = self.matrix_from_tensor(cur_rot, cur_trans) # 轴角/四元素 转 变换矩阵 T。这里是轴角
# ********************* 从当前帧中采样像素点,并从数据中读取这些像素点的观测值 *********************
# Note here we fix the sampled points for optimisation
if indice is None:
# 从一个范围内的整数(0 到 H * W - 1)中随机选择samples=1024个图片像素点索引
indice = self.select_samples(self.dataset.H-iH*2, self.dataset.W-iW*2, self.config['tracking']['sample'])
# Slicing
indice_h, indice_w = indice % (self.dataset.H - iH * 2), indice // (self.dataset.H - iH * 2) # 将取样点的像素坐标(高和宽)提取出来
rays_d_cam = batch['direction'].squeeze(0)[iH:-iH, iW:-iW, :][indice_h, indice_w, :].to(self.device) # 相机坐标下各个采样点的射线方向 [1024, 3]
target_s = batch['rgb'].squeeze(0)[iH:-iH, iW:-iW, :][indice_h, indice_w, :].to(self.device) # 相机坐标下各个采样点的rgb观测值[1024, 3]
target_d = batch['depth'].squeeze(0)[iH:-iH, iW:-iW][indice_h, indice_w].to(self.device).unsqueeze(-1) # 相机坐标下各个采样点的深度观测值 [1024, 1]
# ********************* 根据位姿估计计算射线起点和方向 *********************
rays_o = c2w_est[...,:3, -1].repeat(self.config['tracking']['sample'], 1) # 射线的原点:即估计位姿T中提取的位移t
rays_d = torch.sum(rays_d_cam[..., None, :] * c2w_est[:, :3, :3], -1) # 射线的方向:相机坐标下的射线方向 ✖ 位姿估计C2W 转为世界坐标系下的射线方向
# ********************* 使用encoder/decoder网络,根据观测值和估计值计算损失函数 *********************
ret = self.model.forward(rays_o, rays_d, target_s, target_d) # 得到rgb图,深度图,rgb损失,深度损失,sdf损失,fs损失
loss = self.get_loss_from_ret(ret) # 所有的loss之和
# ********************* 判断sdf损失有无变小,并找到最佳loss下的位姿估计 *********************
if best_sdf_loss is None: # 初始化best_sdf_loss
best_sdf_loss = loss.cpu().item() # 将loss的数据结构之和从tensor转为double
best_c2w_est = c2w_est.detach() # 创建张量c2w_est的无梯度副本
with torch.no_grad(): # 下面代码块:禁用梯度计算
c2w_est = self.matrix_from_tensor(cur_rot, cur_trans) # 将轴角cur_rot 和 位移cur_trans 转为 变换矩阵
if loss.cpu().item() < best_sdf_loss: # 如果损失相比之前变小了,就更新最好的sdf损失
best_sdf_loss = loss.cpu().item() # 更新最好的sdf损失
best_c2w_est = c2w_est.detach() # 创建张量c2w_est的无梯度副本,保存最佳loss下的位姿估计
thresh = 0
else:
thresh +=1 # 如果优化后损失没比以前变小,thresh+1
if thresh >self.config['tracking']['wait_iters']: # wait_iters=100
break
# ********************* 更新参数 *********************
loss.backward() # 反响传播,计算相对于 cur_trans 和 cur_rot 的梯度
pose_optimizer.step() # 更新参数
# ********************* 结束跟踪 *********************
# <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
# ********************* 使用loss最小的位姿估计作为当前帧的位姿估计 *********************
if self.config['tracking']['best']:
# Use the pose with smallest loss
self.est_c2w_data[frame_id] = best_c2w_est.detach().clone()[0]
else:
# Use the pose after the last iteration
self.est_c2w_data[frame_id] = c2w_est.detach().clone()[0]
# ! -------------------- 2.3 Tracked frame --------------------
# Save relative pose of non-keyframes
if frame_id % self.config['mapping']['keyframe_every'] != 0: # 如果不是关键帧
kf_id = frame_id // self.config['mapping']['keyframe_every'] # 前帧所属的关键帧的索引,比如11//5=2 第11帧属于第2个的关键帧
kf_frame_id = kf_id * self.config['mapping']['keyframe_every'] # 关键帧id,比如2*5,第2个关键帧就是第10帧
c2w_key = self.est_c2w_data[kf_frame_id] # 关键帧的估计位姿
delta = self.est_c2w_data[frame_id] @ c2w_key.float().inverse() # 当前帧的T 乘上 关键帧的T^-1 得到两帧位姿之间的差异
self.est_c2w_data_rel[frame_id] = delta # 保存差异
print('Best loss: {}, Last loss{}'.format(F.l1_loss(best_c2w_est.to(self.device)[0,:3], c2w_gt[:3]).cpu().item(), F.l1_loss(c2w_est[0,:3], c2w_gt[:3]).cpu().item()))
def convert_relative_pose(self):
poses = {}
for i in range(len(self.est_c2w_data)):
if i % self.config['mapping']['keyframe_every'] == 0:
poses[i] = self.est_c2w_data[i]
else:
kf_id = i // self.config['mapping']['keyframe_every']
kf_frame_id = kf_id * self.config['mapping']['keyframe_every']
c2w_key = self.est_c2w_data[kf_frame_id]
delta = self.est_c2w_data_rel[i]
poses[i] = delta @ c2w_key
return poses
def create_optimizer(self):
'''
Create optimizer for mapping
'''
# ********************* Optimizer for BA *********************
# params: encoder/decoder网络的参数
trainable_parameters = [{'params': self.model.decoder.parameters(), 'weight_decay': 1e-6, 'lr': self.config['mapping']['lr_decoder']},
{'params': self.model.embed_fn.parameters(), 'eps': 1e-15, 'lr': self.config['mapping']['lr_embed']}]
if not self.config['grid']['oneGrid']:
trainable_parameters.append({'params': self.model.embed_fn_color.parameters(), 'eps': 1e-15, 'lr': self.config['mapping']['lr_embed_color']})
# 创建 Adam 优化器,用于更新模型中的可训练参数(trainable_parameters)
self.map_optimizer = optim.Adam(trainable_parameters, betas=(0.9, 0.99))
#********************* Optimizer for current frame mapping *********************
if self.config['mapping']['cur_frame_iters'] > 0:
params_cur_mapping = [{'params': self.model.embed_fn.parameters(), 'eps': 1e-15, 'lr': self.config['mapping']['lr_embed']}]
if not self.config['grid']['oneGrid']:
params_cur_mapping.append({'params': self.model.embed_fn_color.parameters(), 'eps': 1e-15, 'lr': self.config['mapping']['lr_embed_color']})
self.cur_map_optimizer = optim.Adam(params_cur_mapping, betas=(0.9, 0.99))
def save_mesh(self, i, voxel_size=0.05):
mesh_savepath = os.path.join(self.config['data']['output'], self.config['data']['exp_name'], 'mesh_track{}.ply'.format(i))
if self.config['mesh']['render_color']:
color_func = self.model.render_surface_color
else:
color_func = self.model.query_color
extract_mesh(self.model.query_sdf,
self.config,
self.bounding_box,
color_func=color_func,
marching_cube_bound=self.marching_cube_bound,
voxel_size=voxel_size,
mesh_savepath=mesh_savepath)
def run(self):
# ********************* 创建map和BA的优化器 *********************
# Adam 优化器,用于优化encoder/decoder网络
# 优化位姿的优化器见tracking_render()
self.create_optimizer()
# ********************* 加载数据 *********************
data_loader = DataLoader(self.dataset, num_workers=self.config['data']['num_workers'])
# ! -------------------- 2/3. Start Co-SLAM(tracking + Mapping) --------------------
"""
tqdm类:给迭代过程显示进度条。
enumerate():同时获得可迭代对象的索引和对应元素,所以i是索引,batch是当前批次的数据。
batch: tum数据集
- frame_id: [i] 同索引i
- c2w: [1,4,4]
- rgb: [1,368,496,3]
- depth: [1,368,496]
- direction:[1,368,496,3]
"""
for i, batch in tqdm(enumerate(data_loader)):
# ********************* 可视化rgb和深度图 *********************
if self.config['mesh']['visualisation']:
rgb = cv2.cvtColor(batch["rgb"].squeeze().cpu().numpy(), cv2.COLOR_BGR2RGB) # 将图片的颜色通道从BGR转为RGB [368, 496, 3]
raw_depth = batch["depth"] # 每一像素的深度作为射线深度 [1,368,496]
mask = (raw_depth >= self.config["cam"]["depth_trunc"]).squeeze(0) # 创建一个掩码,用于过滤小于5的深度 [368,496]
depth_colormap = colormap_image(batch["depth"]) # 将1通道的深度图像转为3通道的颜色图 [3, 368, 496]
depth_colormap[:, mask] = 255. # 用掩码将深度小于5的像素的rgb都设为255
depth_colormap = depth_colormap.permute(1, 2, 0).cpu().numpy() # 换一下排列 [368, 496, 3]
image = np.hstack((rgb, depth_colormap)) # 水平方向上合并rgb图和深度图 [368, 992, 3]
cv2.namedWindow('RGB-D'.format(i), cv2.WINDOW_AUTOSIZE) # 在窗口中显示上面这2张图
cv2.imshow('RGB-D'.format(i), image)
key = cv2.waitKey(1)
# ********************* 建立初始的 地图和位姿估计 *********************
# First frame mapping
if i == 0:
self.first_frame_mapping(batch, self.config['mapping']['first_iters']) # ? 包含2.1
# ********************* 建立每一帧的地图和位姿估计 *********************
# Tracking + Mapping
else:
# ! -------------------- 2. tracking --------------------
if self.config['tracking']['iter_point'] > 0:
# 通过点云损失来跟踪当前帧的相机位姿,本论文没用该方法
self.tracking_pc(batch, i)
# 使用当前帧的rgb损失,深度损失,sdf损失,fs损失来跟踪当前帧的相机位姿
self.tracking_render(batch, i) # ? 包含2.1, 2.2, 2.3
# ! -------------------- 3. Mapping --------------------
if i%self.config['mapping']['map_every']==0: # 每5帧建一次图
self.current_frame_mapping(batch, i) # ? 包含3.2
# ! -------------------- 3.3 BA --------------------
self.global_BA(batch, i)
# ! -------------------- 3.1 Keyframe database --------------------
# Add keyframe
if i % self.config['mapping']['keyframe_every'] == 0:
self.keyframeDatabase.add_keyframe(batch, filter_depth=self.config['mapping']['filter_depth'])
print('add keyframe:',i)
# ! -------------------- * Evaluation --------------------
if i % self.config['mesh']['vis']==0:
self.save_mesh(i, voxel_size=self.config['mesh']['voxel_eval'])
pose_relative = self.convert_relative_pose()
pose_evaluation(self.pose_gt, self.est_c2w_data, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i)
pose_evaluation(self.pose_gt, pose_relative, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i, img='pose_r', name='output_relative.txt')
# 展示轨迹真值和预测轨迹图
if cfg['mesh']['visualisation']:
cv2.namedWindow('Traj:'.format(i), cv2.WINDOW_AUTOSIZE)
traj_image = cv2.imread(os.path.join(self.config['data']['output'], self.config['data']['exp_name'], "pose_r_{}.png".format(i)))
# best_traj_image = cv2.imread(os.path.join(best_logdir_scene, "pose_r_{}.png".format(i)))
# image_show = np.hstack((traj_image, best_traj_image))
image_show = traj_image
cv2.imshow('Traj:'.format(i), image_show)
key = cv2.waitKey(1)
model_savepath = os.path.join(self.config['data']['output'], self.config['data']['exp_name'], 'checkpoint{}.pt'.format(i))
self.save_ckpt(model_savepath)
self.save_mesh(i, voxel_size=self.config['mesh']['voxel_final'])
pose_relative = self.convert_relative_pose()
pose_evaluation(self.pose_gt, self.est_c2w_data, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i)
pose_evaluation(self.pose_gt, pose_relative, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i, img='pose_r', name='output_relative.txt')
#TODO: Evaluation of reconstruction
if __name__ == '__main__':
# ********************* 加载参数 *********************
print('Start running...')
parser = argparse.ArgumentParser(
description='Arguments for running the NICE-SLAM/iMAP*.'
)
parser.add_argument('--config', type=str, help='Path to config file.')
parser.add_argument('--input_folder', type=str,
help='input folder, this have higher priority, can overwrite the one in config file')
parser.add_argument('--output', type=str,
help='output folder, this have higher priority, can overwrite the one in config file')
args = parser.parse_args()
cfg = config.load_config(args.config)
if args.output is not None:
cfg['data']['output'] = args.output
print("Saving config and script...")
save_path = os.path.join(cfg["data"]["output"], cfg['data']['exp_name']) # save_path: "output/TUM/fr_desk/demo"
if not os.path.exists(save_path):
os.makedirs(save_path)
shutil.copy("coslam.py", os.path.join(save_path, 'coslam.py'))
with open(os.path.join(save_path, 'config.json'),"w", encoding='utf-8') as f:
f.write(json.dumps(cfg, indent=4)) # 将名为 cfg 的 Python 字典转换为每级缩进4个空格的 JSON 格式字符串。
# ********************* 开始SLAM *********************
# ! -------------------- 1. Scene representation: 网络构建 --------------------
slam = CoSLAM(cfg)
# ! -------------------- 2/3. Start Co-SLAM(tracking + Mapping) --------------------
slam.run()