Newer
Older
#include <ftl/operators/mvmls.hpp>
#include "smoothing_cuda.hpp"
#include <ftl/utility/matrix_conversion.hpp>
#include "mvmls_cuda.hpp"
#include <opencv2/cudaarithm.hpp>
using ftl::operators::MultiViewMLS;
using ftl::codecs::Channel;
using cv::cuda::GpuMat;
using ftl::rgbd::Format;
MultiViewMLS::MultiViewMLS(ftl::Configurable *cfg) : ftl::operators::Operator(cfg) {
}
MultiViewMLS::~MultiViewMLS() {
}
bool MultiViewMLS::apply(ftl::rgbd::FrameSet &in, ftl::rgbd::FrameSet &out, cudaStream_t stream) {
cv::cuda::Stream cvstream = cv::cuda::StreamAccessor::wrapStream(stream);
// float thresh = config()->value("mls_threshold", 0.4f);
float disconPixels = config()->value("discon_pixels", 100.0f); // Max before definitely not same surface
float col_smooth = config()->value("mls_colour_smoothing", 30.0f);
int iters = config()->value("mls_iterations", 3);
int radius = config()->value("mls_radius",5);
//bool aggre = config()->value("aggregation", true);
//int win = config()->value("cost_function",1);
int win = config()->value("window_size",16);
bool do_corr = config()->value("merge_corresponding", true);
bool do_aggr = config()->value("merge_mls", false);
bool do_colour_adjust = config()->value("apply_colour_adjust", false);
bool cull_zero = config()->value("cull_no_confidence", false);
//bool show_best_source = config()->value("show_pixel_source", false);
ftl::cuda::MvMLSParams params;
params.colour_smooth = config()->value("colour_smooth", 150.0f);
params.spatial_smooth = config()->value("spatial_smooth", 1.0f);
params.sub_pixel = config()->value("sub_pixel", 0.5f);
params.P1 = config()->value("P1", 0.1f);
params.P2 = config()->value("P2", 0.2f);
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
bool show_consistency = config()->value("show_consistency", false);
bool show_adjustment = config()->value("show_adjustment", false);
if (in.frames.size() < 1) return false;
auto size = in.frames[0].get<GpuMat>(Channel::Depth).size();
// Make sure we have enough buffers
while (normals_horiz_.size() < in.frames.size()) {
normals_horiz_.push_back(new ftl::cuda::TextureObject<half4>(size.height, size.width));
centroid_horiz_.push_back(new ftl::cuda::TextureObject<float4>(size.height, size.width));
centroid_vert_.push_back(new ftl::cuda::TextureObject<float4>(size.width, size.height));
contributions_.push_back(new ftl::cuda::TextureObject<float>(size.width, size.height));
}
// Make sure all buffers are at correct resolution and are allocated
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f = in.frames[i];
auto size = f.get<GpuMat>(Channel::Depth).size();
centroid_horiz_[i]->create(size.height, size.width);
normals_horiz_[i]->create(size.height, size.width);
centroid_vert_[i]->create(size.width, size.height);
contributions_[i]->create(size.width, size.height);
if (!f.hasChannel(Channel::Normals)) {
throw FTL_Error("Required normals channel missing for MLS");
}
if (!f.hasChannel(Channel::Support2) && !f.hasChannel(Channel::Support1)) {
throw FTL_Error("Required cross support channel missing for MLS");
}
// Create required channels
f.create<GpuMat>(Channel::Confidence, Format<float>(size));
f.createTexture<float>(Channel::Confidence);
f.create<GpuMat>(Channel::Screen, Format<short2>(size));
f.createTexture<short2>(Channel::Screen);
f.get<GpuMat>(Channel::Confidence).setTo(cv::Scalar(0.0f), cvstream);
}
//for (int iter=0; iter<iters; ++iter) {
// Step 1:
// Calculate correspondences and adjust depth values
// Step 2:
// Find corresponding points and perform aggregation of any correspondences
// For each camera combination
if (do_corr) {
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f1 = in.frames[i];
//f1.get<GpuMat>(Channel::Depth2).setTo(cv::Scalar(0.0f), cvstream);
//f1.get<GpuMat>(Channel::Confidence).setTo(cv::Scalar(0.0f), cvstream);
Eigen::Vector4d d1(0.0, 0.0, 1.0, 0.0);
d1 = f1.getPose() * d1;
for (size_t j=i+1; j<in.frames.size(); ++j) {
if (i == j) continue;
//LOG(INFO) << "Running phase1";
auto &f2 = in.frames[j];
//auto s1 = in.sources[i];
//auto s2 = in.sources[j];
// Are cameras facing similar enough direction?
Eigen::Vector4d d2(0.0, 0.0, 1.0, 0.0);
d2 = f2.getPose() * d2;
// No, so skip this combination
if (d1.dot(d2) <= 0.0) continue;
auto pose = MatrixConversion::toCUDA(f1.getPose().cast<float>().inverse() * f2.getPose().cast<float>());
auto pose2 = MatrixConversion::toCUDA(f2.getPose().cast<float>().inverse() * f1.getPose().cast<float>());
//auto transform = pose2 * pose1;
//Calculate screen positions of estimated corresponding points
//if (iter == 0) {
ftl::cuda::correspondence(
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<short2>(Channel::Screen),
f1.getTexture<float>(Channel::Confidence),
f1.getTexture<uint8_t>(Channel::Mask),
pose2,
f1.getLeftCamera(),
f2.getLeftCamera(),
params,
win,
stream
);
ftl::cuda::correspondence(
f2.getTexture<float>(Channel::Depth),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<short2>(Channel::Screen),
f2.getTexture<float>(Channel::Confidence),
f2.getTexture<uint8_t>(Channel::Mask),
pose,
f2.getLeftCamera(),
f1.getLeftCamera(),
params,
win,
stream
);
/*ftl::cuda::remove_cor_error(
f1.getTexture<float>(Channel::Confidence),
f1.getTexture<short2>(Channel::Screen),
f2.getTexture<short2>(Channel::Screen),
2.0f,
stream);
ftl::cuda::remove_cor_error(
f2.getTexture<float>(Channel::Confidence),
f2.getTexture<short2>(Channel::Screen),
f1.getTexture<short2>(Channel::Screen),
2.0f,
stream);*/
// Actually perform the adjustment
cv::cuda::add(f1.get<GpuMat>(Channel::Depth), f1.get<GpuMat>(Channel::Confidence), f1.get<GpuMat>(Channel::Depth), cv::noArray(), -1, cvstream);
cv::cuda::add(f2.get<GpuMat>(Channel::Depth), f2.get<GpuMat>(Channel::Confidence), f2.get<GpuMat>(Channel::Depth), cv::noArray(), -1, cvstream);
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
/*ftl::cuda::viz_reprojection(
f1.getTexture<uchar4>(Channel::Colour),
f2.getTexture<uchar4>(Channel::Colour),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
pose,
f2.getLeftCamera(),
f1.getLeftCamera(),
stream
);
ftl::cuda::viz_reprojection(
f2.getTexture<uchar4>(Channel::Colour),
f1.getTexture<uchar4>(Channel::Colour),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<float>(Channel::Depth),
pose2,
f1.getLeftCamera(),
f2.getLeftCamera(),
stream
);*/
//continue;
//auto &g1 = f1.get<GpuMat>(Channel::Depth);
//costs_[0].create(g1.cols/4, (g1.rows/4) * 16);
//auto &g2 = f2.get<GpuMat>(Channel::Depth);
//costs_[1].create(g2.cols/4, (g2.rows/4) * 16);
/*ftl::cuda::correspondence(
f1.getTexture<uchar4>(Channel::Colour),
f2.getTexture<uchar4>(Channel::Colour),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<short2>(Channel::Screen),
f1.getTexture<float>(Channel::Confidence),
//f1.getTexture<uint8_t>(Channel::Mask),
//costs_[0],
pose2,
f1.getLeftCamera(),
f2.getLeftCamera(),
params,
16,
stream
);
ftl::cuda::correspondence(
f2.getTexture<uchar4>(Channel::Colour),
f1.getTexture<uchar4>(Channel::Colour),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<short2>(Channel::Screen),
f2.getTexture<float>(Channel::Confidence),
//f2.getTexture<uint8_t>(Channel::Mask),
//costs_[1],
pose,
f2.getLeftCamera(),
f1.getLeftCamera(),
params,
16,
stream
);*/
/*ftl::cuda::aggregate_colour_costs(
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
costs_[0],
costs_[1],
f1.getTexture<short2>(Channel::Screen),
f1.getTexture<float>(Channel::Confidence),
pose2,
f1.getLeftCamera(),
f2.getLeftCamera(),
params,
16,
stream
);
ftl::cuda::aggregate_colour_costs(
f2.getTexture<float>(Channel::Depth),
f1.getTexture<float>(Channel::Depth),
costs_[1],
costs_[0],
f2.getTexture<short2>(Channel::Screen),
f2.getTexture<float>(Channel::Confidence),
pose,
f2.getLeftCamera(),
f1.getLeftCamera(),
params,
16,
stream
);*/
if (show_consistency) {
ftl::cuda::show_cor_error(f1.getTexture<uchar4>(Channel::Colour), f1.getTexture<short2>(Channel::Screen), f2.getTexture<short2>(Channel::Screen), 5.0f, stream);
ftl::cuda::show_cor_error(f2.getTexture<uchar4>(Channel::Colour), f2.getTexture<short2>(Channel::Screen), f1.getTexture<short2>(Channel::Screen), 5.0f, stream);
}
/*ftl::cuda::remove_cor_error(
f1.getTexture<float>(Channel::Confidence),
f1.getTexture<short2>(Channel::Screen),
f2.getTexture<short2>(Channel::Screen),
5.0f,
stream);
ftl::cuda::remove_cor_error(
f2.getTexture<float>(Channel::Confidence),
f2.getTexture<short2>(Channel::Screen),
f1.getTexture<short2>(Channel::Screen),
5.0f,
stream);*/
// Actually perform the colour adjustment
//if (do_colour_adjust) {
// cv::cuda::add(f1.get<GpuMat>(Channel::Depth), f1.get<GpuMat>(Channel::Confidence), f1.get<GpuMat>(Channel::Depth), cv::noArray(), -1, cvstream);
// cv::cuda::add(f2.get<GpuMat>(Channel::Depth), f2.get<GpuMat>(Channel::Confidence), f2.get<GpuMat>(Channel::Depth), cv::noArray(), -1, cvstream);
//}
if (show_adjustment) {
ftl::cuda::show_depth_adjustment(f1.getTexture<uchar4>(Channel::Colour), f1.getTexture<short2>(Channel::Screen), f1.getTexture<float>(Channel::Confidence), 0.04f, stream);
ftl::cuda::show_depth_adjustment(f2.getTexture<uchar4>(Channel::Colour), f2.getTexture<short2>(Channel::Screen), f2.getTexture<float>(Channel::Confidence), 0.04f, stream);
}
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
//} //else {
/*ftl::cuda::correspondence(
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<uchar4>(Channel::Colour),
f2.getTexture<uchar4>(Channel::Colour),
// TODO: Add normals and other things...
f1.getTexture<short2>(Channel::Screen),
f1.getTexture<float>(Channel::Confidence),
f1.getTexture<uint8_t>(Channel::Mask),
pose2,
f1.getLeftCamera(),
f2.getLeftCamera(),
params,
win,
stream
);
ftl::cuda::correspondence(
f2.getTexture<float>(Channel::Depth),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<uchar4>(Channel::Colour),
f1.getTexture<uchar4>(Channel::Colour),
// TODO: Add normals and other things...
f2.getTexture<short2>(Channel::Screen),
f2.getTexture<float>(Channel::Confidence),
f2.getTexture<uint8_t>(Channel::Mask),
pose,
f2.getLeftCamera(),
f1.getLeftCamera(),
params,
win,
stream
);*/
//}
// Also calculate best source for each point
/*ftl::cuda::best_sources(
f1.getTexture<uchar4>(Channel::Support1),
f2.getTexture<uchar4>(Channel::Support1),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<short2>(Channel::Screen),
transform,
s1->parameters(),
s2->parameters(),
i, j, stream
);*/
}
}
// Reduce window size for next iteration
//win = max(win>>1, 4);
// Redo normals
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f = in.frames[i];
ftl::cuda::normals(
f.getTexture<half4>(Channel::Normals),
f.getTexture<float>(Channel::Depth),
f.getLeftCamera(), stream
);
}
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
}
for (int iter=0; iter<iters; ++iter) {
// Find best source for every pixel
/*for (size_t i=0; i<in.frames.size(); ++i) {
auto &f1 = in.frames[i];
//f1.get<GpuMat>(Channel::Depth2).setTo(cv::Scalar(0.0f), cvstream);
//f1.get<GpuMat>(Channel::Confidence).setTo(cv::Scalar(0.0f), cvstream);
Eigen::Vector4d d1(0.0, 0.0, 1.0, 0.0);
d1 = in.sources[i]->getPose() * d1;
for (size_t j=0; j<in.frames.size(); ++j) {
if (i == j) continue;
//LOG(INFO) << "Running phase1";
auto &f2 = in.frames[j];
auto s1 = in.sources[i];
auto s2 = in.sources[j];
// Are cameras facing similar enough direction?
Eigen::Vector4d d2(0.0, 0.0, 1.0, 0.0);
d2 = in.sources[j]->getPose() * d2;
// No, so skip this combination
if (d1.dot(d2) <= 0.0) continue;
auto pose1 = MatrixConversion::toCUDA(s1->getPose().cast<float>());
auto pose1_inv = MatrixConversion::toCUDA(s1->getPose().cast<float>().inverse());
auto pose2 = MatrixConversion::toCUDA(s2->getPose().cast<float>().inverse());
auto pose2_inv = MatrixConversion::toCUDA(s2->getPose().cast<float>());
auto transform = pose2 * pose1;
// Also calculate best source for each point
ftl::cuda::best_sources(
f1.getTexture<float4>(Channel::Normals),
f2.getTexture<float4>(Channel::Normals),
f1.getTexture<uchar4>(Channel::Support1),
f2.getTexture<uchar4>(Channel::Support1),
f1.getTexture<float>(Channel::Depth),
f2.getTexture<float>(Channel::Depth),
f1.getTexture<short2>(Channel::Screen),
transform,
s1->parameters(),
s2->parameters(),
i, j, stream
);
}
}*/
if (radius > 0) {
// Step 2:
// Do the horizontal and vertical MLS aggregations for each source
// But don't do the final move step.
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f = in.frames[i];
//auto *s = in.sources[i];
// Clear data
//cv::cuda::GpuMat data(contributions_[i]->height(), contributions_[i]->width(), CV_32F, contributions_[i]->devicePtr(), contributions_[i]->pixelPitch());
//data.setTo(cv::Scalar(0.0f), cvstream);
if (cull_zero && iter == iters-1) {
ftl::cuda::zero_confidence(
f.getTexture<float>(Channel::Confidence),
f.getTexture<float>(Channel::Depth),
stream
);
}
float thresh = (1.0f / f.getLeft().fx) * disconPixels;
ftl::cuda::mls_aggr_horiz(
f.createTexture<uchar4>((f.hasChannel(Channel::Support2)) ? Channel::Support2 : Channel::Support1),
f.createTexture<half4>(Channel::Normals),
*normals_horiz_[i],
f.createTexture<float>(Channel::Depth),
*centroid_horiz_[i],
f.createTexture<uchar4>(Channel::Colour),
thresh,
col_smooth,
radius,
f.getLeftCamera(),
stream
);
ftl::cuda::mls_aggr_vert(
f.getTexture<uchar4>((f.hasChannel(Channel::Support2)) ? Channel::Support2 : Channel::Support1),
*normals_horiz_[i],
f.getTexture<half4>(Channel::Normals),
*centroid_horiz_[i],
*centroid_vert_[i],
thresh,
col_smooth,
radius,
f.getLeftCamera(),
stream
);
}
//return true;
// Step 3:
// Find corresponding points and perform aggregation of any correspondences
// For each camera combination
if (do_aggr) {
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f1 = in.frames[i];
//f1.get<GpuMat>(Channel::Depth2).setTo(cv::Scalar(0.0f), cvstream);
//f1.get<GpuMat>(Channel::Confidence).setTo(cv::Scalar(0.0f), cvstream);
Eigen::Vector4d d1(0.0, 0.0, 1.0, 0.0);
d1 = f1.getPose() * d1;
for (size_t j=0; j<in.frames.size(); ++j) {
if (i == j) continue;
//LOG(INFO) << "Running phase1";
auto &f2 = in.frames[j];
//auto s1 = in.sources[i];
//auto s2 = in.sources[j];
// Are cameras facing similar enough direction?
Eigen::Vector4d d2(0.0, 0.0, 1.0, 0.0);
d2 = f2.getPose() * d2;
// No, so skip this combination
if (d1.dot(d2) <= 0.0) continue;
auto pose1 = MatrixConversion::toCUDA(f1.getPose().cast<float>());
//auto pose1_inv = MatrixConversion::toCUDA(f1.getPose().cast<float>().inverse());
auto pose2 = MatrixConversion::toCUDA(f2.getPose().cast<float>().inverse());
//auto pose2_inv = MatrixConversion::toCUDA(f2.getPose().cast<float>());
auto transform = pose2 * pose1;
// For the corresponding points, combine normals and centroids
ftl::cuda::aggregate_sources(
f1.getTexture<half4>(Channel::Normals),
f2.getTexture<half4>(Channel::Normals),
*centroid_vert_[i],
*centroid_vert_[j],
f1.getTexture<float>(Channel::Depth),
//contributions_[i],
//contributions_[j],
//f1.getTexture<short2>(Channel::Screen),
transform,
f1.getLeftCamera(),
f2.getLeftCamera(),
stream
);
//LOG(INFO) << "Correspondences done... " << i;
}
}
}
// Step 3:
// Normalise aggregations and move the points
for (size_t i=0; i<in.frames.size(); ++i) {
auto &f = in.frames[i];
//auto *s = in.sources[i];
auto size = f.get<GpuMat>(Channel::Depth).size();
/*if (do_corr) {
ftl::cuda::normalise_aggregations(
f.getTexture<float4>(Channel::Normals),
centroid_vert_[i],
contributions_[i],
stream
);
}*/
ftl::cuda::mls_adjust_depth(
f.getTexture<half4>(Channel::Normals),
*centroid_vert_[i],
f.createTexture<float>(Channel::Depth2, ftl::rgbd::Format<float>(size)),
f.getLeftCamera(),
stream
);
f.swapChannels(Channel::Depth, Channel::Depth2);
/*if (show_best_source) {
ftl::cuda::vis_best_sources(
f.getTexture<short2>(Channel::Screen),
f.getTexture<uchar4>(Channel::Colour),
i,
7, stream
);
}*/
}
}
}
return true;
}