-
Notifications
You must be signed in to change notification settings - Fork 93
/
module.py
631 lines (488 loc) · 24.6 KB
/
module.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
import sys
sys.path.append("..")
from utils import local_pcd
def init_bn(module):
if module.weight is not None:
nn.init.ones_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
return
def init_uniform(module, init_method):
if module.weight is not None:
if init_method == "kaiming":
nn.init.kaiming_uniform_(module.weight)
elif init_method == "xavier":
nn.init.xavier_uniform_(module.weight)
return
class Conv2d(nn.Module):
"""Applies a 2D convolution (optionally with batch normalization and relu activation)
over an input signal composed of several input planes.
Attributes:
conv (nn.Module): convolution module
bn (nn.Module): batch normalization module
relu (bool): whether to activate by relu
Notes:
Default momentum for batch normalization is set to be 0.01,
"""
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
relu=True, bn=True, bn_momentum=0.1, init_method="xavier", **kwargs):
super(Conv2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride=stride,
bias=(not bn), **kwargs)
self.kernel_size = kernel_size
self.stride = stride
self.bn = nn.BatchNorm2d(out_channels, momentum=bn_momentum) if bn else None
self.relu = relu
# assert init_method in ["kaiming", "xavier"]
# self.init_weights(init_method)
def forward(self, x):
x = self.conv(x)
if self.bn is not None:
x = self.bn(x)
if self.relu:
x = F.relu(x, inplace=True)
return x
def init_weights(self, init_method):
"""default initialization"""
init_uniform(self.conv, init_method)
if self.bn is not None:
init_bn(self.bn)
class Deconv2d(nn.Module):
"""Applies a 2D deconvolution (optionally with batch normalization and relu activation)
over an input signal composed of several input planes.
Attributes:
conv (nn.Module): convolution module
bn (nn.Module): batch normalization module
relu (bool): whether to activate by relu
Notes:
Default momentum for batch normalization is set to be 0.01,
"""
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
relu=True, bn=True, bn_momentum=0.1, init_method="xavier", **kwargs):
super(Deconv2d, self).__init__()
self.out_channels = out_channels
assert stride in [1, 2]
self.stride = stride
self.conv = nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=stride,
bias=(not bn), **kwargs)
self.bn = nn.BatchNorm2d(out_channels, momentum=bn_momentum) if bn else None
self.relu = relu
# assert init_method in ["kaiming", "xavier"]
# self.init_weights(init_method)
def forward(self, x):
y = self.conv(x)
if self.stride == 2:
h, w = list(x.size())[2:]
y = y[:, :, :2 * h, :2 * w].contiguous()
if self.bn is not None:
x = self.bn(y)
if self.relu:
x = F.relu(x, inplace=True)
return x
def init_weights(self, init_method):
"""default initialization"""
init_uniform(self.conv, init_method)
if self.bn is not None:
init_bn(self.bn)
class Conv3d(nn.Module):
"""Applies a 3D convolution (optionally with batch normalization and relu activation)
over an input signal composed of several input planes.
Attributes:
conv (nn.Module): convolution module
bn (nn.Module): batch normalization module
relu (bool): whether to activate by relu
Notes:
Default momentum for batch normalization is set to be 0.01,
"""
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1,
relu=True, bn=True, bn_momentum=0.1, init_method="xavier", **kwargs):
super(Conv3d, self).__init__()
self.out_channels = out_channels
self.kernel_size = kernel_size
assert stride in [1, 2]
self.stride = stride
self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride,
bias=(not bn), **kwargs)
self.bn = nn.BatchNorm3d(out_channels, momentum=bn_momentum) if bn else None
self.relu = relu
# assert init_method in ["kaiming", "xavier"]
# self.init_weights(init_method)
def forward(self, x):
x = self.conv(x)
if self.bn is not None:
x = self.bn(x)
if self.relu:
x = F.relu(x, inplace=True)
return x
def init_weights(self, init_method):
"""default initialization"""
init_uniform(self.conv, init_method)
if self.bn is not None:
init_bn(self.bn)
class Deconv3d(nn.Module):
"""Applies a 3D deconvolution (optionally with batch normalization and relu activation)
over an input signal composed of several input planes.
Attributes:
conv (nn.Module): convolution module
bn (nn.Module): batch normalization module
relu (bool): whether to activate by relu
Notes:
Default momentum for batch normalization is set to be 0.01,
"""
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1,
relu=True, bn=True, bn_momentum=0.1, init_method="xavier", **kwargs):
super(Deconv3d, self).__init__()
self.out_channels = out_channels
assert stride in [1, 2]
self.stride = stride
self.conv = nn.ConvTranspose3d(in_channels, out_channels, kernel_size, stride=stride,
bias=(not bn), **kwargs)
self.bn = nn.BatchNorm3d(out_channels, momentum=bn_momentum) if bn else None
self.relu = relu
# assert init_method in ["kaiming", "xavier"]
# self.init_weights(init_method)
def forward(self, x):
y = self.conv(x)
if self.bn is not None:
x = self.bn(y)
if self.relu:
x = F.relu(x, inplace=True)
return x
def init_weights(self, init_method):
"""default initialization"""
init_uniform(self.conv, init_method)
if self.bn is not None:
init_bn(self.bn)
class ConvBnReLU(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, pad=1):
super(ConvBnReLU, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, bias=False)
self.bn = nn.BatchNorm2d(out_channels)
def forward(self, x):
return F.relu(self.bn(self.conv(x)), inplace=True)
class ConvBn(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, pad=1):
super(ConvBn, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, bias=False)
self.bn = nn.BatchNorm2d(out_channels)
def forward(self, x):
return self.bn(self.conv(x))
class ConvBnReLU3D(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, pad=1):
super(ConvBnReLU3D, self).__init__()
self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, bias=False)
self.bn = nn.BatchNorm3d(out_channels)
def forward(self, x):
return F.relu(self.bn(self.conv(x)), inplace=True)
class ConvBn3D(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, pad=1):
super(ConvBn3D, self).__init__()
self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, bias=False)
self.bn = nn.BatchNorm3d(out_channels)
def forward(self, x):
return self.bn(self.conv(x))
class BasicBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = ConvBnReLU(in_channels, out_channels, kernel_size=3, stride=stride, pad=1)
self.conv2 = ConvBn(out_channels, out_channels, kernel_size=3, stride=1, pad=1)
self.downsample = downsample
self.stride = stride
def forward(self, x):
out = self.conv1(x)
out = self.conv2(out)
if self.downsample is not None:
x = self.downsample(x)
out += x
return out
class Hourglass3d(nn.Module):
def __init__(self, channels):
super(Hourglass3d, self).__init__()
self.conv1a = ConvBnReLU3D(channels, channels * 2, kernel_size=3, stride=2, pad=1)
self.conv1b = ConvBnReLU3D(channels * 2, channels * 2, kernel_size=3, stride=1, pad=1)
self.conv2a = ConvBnReLU3D(channels * 2, channels * 4, kernel_size=3, stride=2, pad=1)
self.conv2b = ConvBnReLU3D(channels * 4, channels * 4, kernel_size=3, stride=1, pad=1)
self.dconv2 = nn.Sequential(
nn.ConvTranspose3d(channels * 4, channels * 2, kernel_size=3, padding=1, output_padding=1, stride=2,
bias=False),
nn.BatchNorm3d(channels * 2))
self.dconv1 = nn.Sequential(
nn.ConvTranspose3d(channels * 2, channels, kernel_size=3, padding=1, output_padding=1, stride=2,
bias=False),
nn.BatchNorm3d(channels))
self.redir1 = ConvBn3D(channels, channels, kernel_size=1, stride=1, pad=0)
self.redir2 = ConvBn3D(channels * 2, channels * 2, kernel_size=1, stride=1, pad=0)
def forward(self, x):
conv1 = self.conv1b(self.conv1a(x))
conv2 = self.conv2b(self.conv2a(conv1))
dconv2 = F.relu(self.dconv2(conv2) + self.redir2(conv1), inplace=True)
dconv1 = F.relu(self.dconv1(dconv2) + self.redir1(x), inplace=True)
return dconv1
def homo_warping(src_fea, src_proj, ref_proj, depth_values):
# src_fea: [B, C, H, W]
# src_proj: [B, 4, 4]
# ref_proj: [B, 4, 4]
# depth_values: [B, Ndepth] o [B, Ndepth, H, W]
# out: [B, C, Ndepth, H, W]
batch, channels = src_fea.shape[0], src_fea.shape[1]
num_depth = depth_values.shape[1]
height, width = src_fea.shape[2], src_fea.shape[3]
with torch.no_grad():
proj = torch.matmul(src_proj, torch.inverse(ref_proj))
rot = proj[:, :3, :3] # [B,3,3]
trans = proj[:, :3, 3:4] # [B,3,1]
y, x = torch.meshgrid([torch.arange(0, height, dtype=torch.float32, device=src_fea.device),
torch.arange(0, width, dtype=torch.float32, device=src_fea.device)])
y, x = y.contiguous(), x.contiguous()
y, x = y.view(height * width), x.view(height * width)
xyz = torch.stack((x, y, torch.ones_like(x))) # [3, H*W]
xyz = torch.unsqueeze(xyz, 0).repeat(batch, 1, 1) # [B, 3, H*W]
rot_xyz = torch.matmul(rot, xyz) # [B, 3, H*W]
rot_depth_xyz = rot_xyz.unsqueeze(2).repeat(1, 1, num_depth, 1) * depth_values.view(batch, 1, num_depth,
-1) # [B, 3, Ndepth, H*W]
proj_xyz = rot_depth_xyz + trans.view(batch, 3, 1, 1) # [B, 3, Ndepth, H*W]
proj_xy = proj_xyz[:, :2, :, :] / proj_xyz[:, 2:3, :, :] # [B, 2, Ndepth, H*W]
proj_x_normalized = proj_xy[:, 0, :, :] / ((width - 1) / 2) - 1
proj_y_normalized = proj_xy[:, 1, :, :] / ((height - 1) / 2) - 1
proj_xy = torch.stack((proj_x_normalized, proj_y_normalized), dim=3) # [B, Ndepth, H*W, 2]
grid = proj_xy
warped_src_fea = F.grid_sample(src_fea, grid.view(batch, num_depth * height, width, 2), mode='bilinear',
padding_mode='zeros')
warped_src_fea = warped_src_fea.view(batch, channels, num_depth, height, width)
return warped_src_fea
class DeConv2dFuse(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, relu=True, bn=True,
bn_momentum=0.1):
super(DeConv2dFuse, self).__init__()
self.deconv = Deconv2d(in_channels, out_channels, kernel_size, stride=2, padding=1, output_padding=1,
bn=True, relu=relu, bn_momentum=bn_momentum)
self.conv = Conv2d(2*out_channels, out_channels, kernel_size, stride=1, padding=1,
bn=bn, relu=relu, bn_momentum=bn_momentum)
# assert init_method in ["kaiming", "xavier"]
# self.init_weights(init_method)
def forward(self, x_pre, x):
x = self.deconv(x)
x = torch.cat((x, x_pre), dim=1)
x = self.conv(x)
return x
class FeatureNet(nn.Module):
def __init__(self, base_channels, num_stage=3, stride=4, arch_mode="unet"):
super(FeatureNet, self).__init__()
assert arch_mode in ["unet", "fpn"], print("mode must be in 'unet' or 'fpn', but get:{}".format(arch_mode))
print("*************feature extraction arch mode:{}****************".format(arch_mode))
self.arch_mode = arch_mode
self.stride = stride
self.base_channels = base_channels
self.num_stage = num_stage
self.conv0 = nn.Sequential(
Conv2d(3, base_channels, 3, 1, padding=1),
Conv2d(base_channels, base_channels, 3, 1, padding=1),
)
self.conv1 = nn.Sequential(
Conv2d(base_channels, base_channels * 2, 5, stride=2, padding=2),
Conv2d(base_channels * 2, base_channels * 2, 3, 1, padding=1),
Conv2d(base_channels * 2, base_channels * 2, 3, 1, padding=1),
)
self.conv2 = nn.Sequential(
Conv2d(base_channels * 2, base_channels * 4, 5, stride=2, padding=2),
Conv2d(base_channels * 4, base_channels * 4, 3, 1, padding=1),
Conv2d(base_channels * 4, base_channels * 4, 3, 1, padding=1),
)
self.out1 = nn.Conv2d(base_channels * 4, base_channels * 4, 1, bias=False)
self.out_channels = [4 * base_channels]
if self.arch_mode == 'unet':
if num_stage == 3:
self.deconv1 = DeConv2dFuse(base_channels * 4, base_channels * 2, 3)
self.deconv2 = DeConv2dFuse(base_channels * 2, base_channels, 3)
self.out2 = nn.Conv2d(base_channels * 2, base_channels * 2, 1, bias=False)
self.out3 = nn.Conv2d(base_channels, base_channels, 1, bias=False)
self.out_channels.append(2 * base_channels)
self.out_channels.append(base_channels)
elif num_stage == 2:
self.deconv1 = DeConv2dFuse(base_channels * 4, base_channels * 2, 3)
self.out2 = nn.Conv2d(base_channels * 2, base_channels * 2, 1, bias=False)
self.out_channels.append(2 * base_channels)
elif self.arch_mode == "fpn":
final_chs = base_channels * 4
if num_stage == 3:
self.inner1 = nn.Conv2d(base_channels * 2, final_chs, 1, bias=True)
self.inner2 = nn.Conv2d(base_channels * 1, final_chs, 1, bias=True)
self.out2 = nn.Conv2d(final_chs, base_channels * 2, 3, padding=1, bias=False)
self.out3 = nn.Conv2d(final_chs, base_channels, 3, padding=1, bias=False)
self.out_channels.append(base_channels * 2)
self.out_channels.append(base_channels)
elif num_stage == 2:
self.inner1 = nn.Conv2d(base_channels * 2, final_chs, 1, bias=True)
self.out2 = nn.Conv2d(final_chs, base_channels, 3, padding=1, bias=False)
self.out_channels.append(base_channels)
def forward(self, x):
conv0 = self.conv0(x)
conv1 = self.conv1(conv0)
conv2 = self.conv2(conv1)
intra_feat = conv2
outputs = {}
out = self.out1(intra_feat)
outputs["stage1"] = out
if self.arch_mode == "unet":
if self.num_stage == 3:
intra_feat = self.deconv1(conv1, intra_feat)
out = self.out2(intra_feat)
outputs["stage2"] = out
intra_feat = self.deconv2(conv0, intra_feat)
out = self.out3(intra_feat)
outputs["stage3"] = out
elif self.num_stage == 2:
intra_feat = self.deconv1(conv1, intra_feat)
out = self.out2(intra_feat)
outputs["stage2"] = out
elif self.arch_mode == "fpn":
if self.num_stage == 3:
intra_feat = F.interpolate(intra_feat, scale_factor=2, mode="nearest") + self.inner1(conv1)
out = self.out2(intra_feat)
outputs["stage2"] = out
intra_feat = F.interpolate(intra_feat, scale_factor=2, mode="nearest") + self.inner2(conv0)
out = self.out3(intra_feat)
outputs["stage3"] = out
elif self.num_stage == 2:
intra_feat = F.interpolate(intra_feat, scale_factor=2, mode="nearest") + self.inner1(conv1)
out = self.out2(intra_feat)
outputs["stage2"] = out
return outputs
class CostRegNet(nn.Module):
def __init__(self, in_channels, base_channels):
super(CostRegNet, self).__init__()
self.conv0 = Conv3d(in_channels, base_channels, padding=1)
self.conv1 = Conv3d(base_channels, base_channels * 2, stride=2, padding=1)
self.conv2 = Conv3d(base_channels * 2, base_channels * 2, padding=1)
self.conv3 = Conv3d(base_channels * 2, base_channels * 4, stride=2, padding=1)
self.conv4 = Conv3d(base_channels * 4, base_channels * 4, padding=1)
self.conv5 = Conv3d(base_channels * 4, base_channels * 8, stride=2, padding=1)
self.conv6 = Conv3d(base_channels * 8, base_channels * 8, padding=1)
self.conv7 = Deconv3d(base_channels * 8, base_channels * 4, stride=2, padding=1, output_padding=1)
self.conv9 = Deconv3d(base_channels * 4, base_channels * 2, stride=2, padding=1, output_padding=1)
self.conv11 = Deconv3d(base_channels * 2, base_channels * 1, stride=2, padding=1, output_padding=1)
self.prob = nn.Conv3d(base_channels, 1, 3, stride=1, padding=1, bias=False)
def forward(self, x):
conv0 = self.conv0(x)
conv2 = self.conv2(self.conv1(conv0))
conv4 = self.conv4(self.conv3(conv2))
x = self.conv6(self.conv5(conv4))
x = conv4 + self.conv7(x)
x = conv2 + self.conv9(x)
x = conv0 + self.conv11(x)
x = self.prob(x)
return x
class RefineNet(nn.Module):
def __init__(self):
super(RefineNet, self).__init__()
self.conv1 = ConvBnReLU(4, 32)
self.conv2 = ConvBnReLU(32, 32)
self.conv3 = ConvBnReLU(32, 32)
self.res = ConvBnReLU(32, 1)
def forward(self, img, depth_init):
concat = F.cat((img, depth_init), dim=1)
depth_residual = self.res(self.conv3(self.conv2(self.conv1(concat))))
depth_refined = depth_init + depth_residual
return depth_refined
def depth_regression(p, depth_values):
if depth_values.dim() <= 2:
# print("regression dim <= 2")
depth_values = depth_values.view(*depth_values.shape, 1, 1)
depth = torch.sum(p * depth_values, 1)
return depth
def cas_mvsnet_loss(inputs, depth_gt_ms, mask_ms, **kwargs):
depth_loss_weights = kwargs.get("dlossw", None)
total_loss = torch.tensor(0.0, dtype=torch.float32, device=mask_ms["stage1"].device, requires_grad=False)
for (stage_inputs, stage_key) in [(inputs[k], k) for k in inputs.keys() if "stage" in k]:
depth_est = stage_inputs["depth"]
depth_gt = depth_gt_ms[stage_key]
mask = mask_ms[stage_key]
mask = mask > 0.5
depth_loss = F.smooth_l1_loss(depth_est[mask], depth_gt[mask], reduction='mean')
if depth_loss_weights is not None:
stage_idx = int(stage_key.replace("stage", "")) - 1
total_loss += depth_loss_weights[stage_idx] * depth_loss
else:
total_loss += 1.0 * depth_loss
return total_loss, depth_loss
def get_cur_depth_range_samples(cur_depth, ndepth, depth_inteval_pixel, shape, max_depth=192.0, min_depth=0.0):
#shape, (B, H, W)
#cur_depth: (B, H, W)
#return depth_range_values: (B, D, H, W)
cur_depth_min = (cur_depth - ndepth / 2 * depth_inteval_pixel) # (B, H, W)
cur_depth_max = (cur_depth + ndepth / 2 * depth_inteval_pixel)
# cur_depth_min = (cur_depth - ndepth / 2 * depth_inteval_pixel).clamp(min=0.0) #(B, H, W)
# cur_depth_max = (cur_depth_min + (ndepth - 1) * depth_inteval_pixel).clamp(max=max_depth)
assert cur_depth.shape == torch.Size(shape), "cur_depth:{}, input shape:{}".format(cur_depth.shape, shape)
new_interval = (cur_depth_max - cur_depth_min) / (ndepth - 1) # (B, H, W)
depth_range_samples = cur_depth_min.unsqueeze(1) + (torch.arange(0, ndepth, device=cur_depth.device,
dtype=cur_depth.dtype,
requires_grad=False).reshape(1, -1, 1,
1) * new_interval.unsqueeze(1))
return depth_range_samples
def get_depth_range_samples(cur_depth, ndepth, depth_inteval_pixel, device, dtype, shape,
max_depth=192.0, min_depth=0.0):
#shape: (B, H, W)
#cur_depth: (B, H, W) or (B, D)
#return depth_range_samples: (B, D, H, W)
if cur_depth.dim() == 2:
cur_depth_min = cur_depth[:, 0] # (B,)
cur_depth_max = cur_depth[:, -1]
new_interval = (cur_depth_max - cur_depth_min) / (ndepth - 1) # (B, )
depth_range_samples = cur_depth_min.unsqueeze(1) + (torch.arange(0, ndepth, device=device, dtype=dtype,
requires_grad=False).reshape(1, -1) * new_interval.unsqueeze(1)) #(B, D)
depth_range_samples = depth_range_samples.unsqueeze(-1).unsqueeze(-1).repeat(1, 1, shape[1], shape[2]) #(B, D, H, W)
else:
depth_range_samples = get_cur_depth_range_samples(cur_depth, ndepth, depth_inteval_pixel, shape, max_depth, min_depth)
return depth_range_samples
if __name__ == "__main__":
# some testing code, just IGNORE it
import sys
sys.path.append("../")
from datasets import find_dataset_def
from torch.utils.data import DataLoader
import numpy as np
import cv2
import matplotlib as mpl
mpl.use('Agg')
import matplotlib.pyplot as plt
# MVSDataset = find_dataset_def("colmap")
# dataset = MVSDataset("../data/results/ford/num10_1/", 3, 'test',
# 128, interval_scale=1.06, max_h=1250, max_w=1024)
MVSDataset = find_dataset_def("dtu_yao")
num_depth = 48
dataset = MVSDataset("../data/DTU/mvs_training/dtu/", '../lists/dtu/train.txt', 'train',
3, num_depth, interval_scale=1.06 * 192 / num_depth)
dataloader = DataLoader(dataset, batch_size=1)
item = next(iter(dataloader))
imgs = item["imgs"][:, :, :, ::4, ::4] #(B, N, 3, H, W)
# imgs = item["imgs"][:, :, :, :, :]
proj_matrices = item["proj_matrices"] #(B, N, 2, 4, 4) dim=N: N view; dim=2: index 0 for extr, 1 for intric
proj_matrices[:, :, 1, :2, :] = proj_matrices[:, :, 1, :2, :]
# proj_matrices[:, :, 1, :2, :] = proj_matrices[:, :, 1, :2, :] * 4
depth_values = item["depth_values"] #(B, D)
imgs = torch.unbind(imgs, 1)
proj_matrices = torch.unbind(proj_matrices, 1)
ref_img, src_imgs = imgs[0], imgs[1:]
ref_proj, src_proj = proj_matrices[0], proj_matrices[1:][0] #only vis first view
src_proj_new = src_proj[:, 0].clone()
src_proj_new[:, :3, :4] = torch.matmul(src_proj[:, 1, :3, :3], src_proj[:, 0, :3, :4])
ref_proj_new = ref_proj[:, 0].clone()
ref_proj_new[:, :3, :4] = torch.matmul(ref_proj[:, 1, :3, :3], ref_proj[:, 0, :3, :4])
warped_imgs = homo_warping(src_imgs[0], src_proj_new, ref_proj_new, depth_values)
ref_img_np = ref_img.permute([0, 2, 3, 1])[0].detach().cpu().numpy()[:, :, ::-1] * 255
cv2.imwrite('../tmp/ref.png', ref_img_np)
cv2.imwrite('../tmp/src.png', src_imgs[0].permute([0, 2, 3, 1])[0].detach().cpu().numpy()[:, :, ::-1] * 255)
for i in range(warped_imgs.shape[2]):
warped_img = warped_imgs[:, :, i, :, :].permute([0, 2, 3, 1]).contiguous()
img_np = warped_img[0].detach().cpu().numpy()
img_np = img_np[:, :, ::-1] * 255
alpha = 0.5
beta = 1 - alpha
gamma = 0
img_add = cv2.addWeighted(ref_img_np, alpha, img_np, beta, gamma)
cv2.imwrite('../tmp/tmp{}.png'.format(i), np.hstack([ref_img_np, img_np, img_add])) #* ratio + img_np*(1-ratio)]))