-
Notifications
You must be signed in to change notification settings - Fork 0
/
coarse_to_fine_patchmatch.cpp
541 lines (452 loc) · 13.9 KB
/
coarse_to_fine_patchmatch.cpp
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
#include "coarse_to_fine_patchmatch.h"
#include <opencv2/xfeatures2d.hpp>
#ifdef WITH_SSE
#include <emmintrin.h>
#endif
static const int NUM_NEIGHBORS = 8;
static const int NEIGHBOR_DX[NUM_NEIGHBORS] = { 0, 0, 1, -1, -1, -1, 1, 1 };
static const int NEIGHBOR_DY[NUM_NEIGHBORS] = { -1, 1, 0, 0, -1, 1, -1, 1 };
template <class T>
static inline const T& clamp(const T& v, const T& lo, const T& hi)
{
return std::max(lo, std::min(v, hi));
}
static std::vector<double> makeScales(double scaleStep, int nscales)
{
std::vector<double> scales(nscales);
double scale = 1;
for (int i = 0; i < nscales; i++)
{
scales[i] = scale;
scale *= scaleStep;
}
return scales;
}
static void GaussianPyrDown(const cv::Mat& src, cv::Mat& dst, double sigma, double scale)
{
cv::Mat tmp;
cv::GaussianBlur(src, tmp, cv::Size(), sigma);
cv::resize(tmp, dst, cv::Size(), scale, scale);
}
static void constructPyramid(const cv::Mat& img, std::vector<cv::Mat>& pyramid, float ratio, int minWidth)
{
// the ratio cannot be arbitrary numbers
if (ratio > 0.98f || ratio < 0.4f)
ratio = 0.75f;
// first decide how many levels
const int nscales = static_cast<int>(log(1. * minWidth / img.cols) / log(ratio));
const std::vector<double> scales = makeScales(ratio, nscales);
pyramid.resize(nscales);
img.copyTo(pyramid[0]);
const double sigma0 = (1 / ratio - 1);
const int n = static_cast<int>(log(0.25) / log(ratio));
for (int s = 1; s < nscales; s++)
{
const double sigma = s <= n ? s * sigma0 : n * sigma0;
const double scale = s <= n ? scales[s] : scales[n];
const cv::Mat& src = s <= n ? img : pyramid[s - n];
GaussianPyrDown(src, pyramid[s], sigma, scale);
}
}
static int makeSeedsAndNeighbors(int w, int h, int step, cv::Mat2f& seeds, cv::Mat1i& neighbors)
{
const int gridw = w / step;
const int gridh = h / step;
const int nseeds = gridw * gridh;
const int ofsx = (w - (gridw - 1) * step) / 2;
const int ofsy = (h - (gridh - 1) * step) / 2;
seeds.create(nseeds, 1);
neighbors.create(nseeds, NUM_NEIGHBORS);
neighbors = -1;
for (int i = 0; i < nseeds; i++)
{
const int x = i % gridw;
const int y = i / gridw;
const float seedx = static_cast<float>(x * step + ofsx);
const float seedy = static_cast<float>(y * step + ofsy);
seeds(i) = cv::Vec2f(seedx, seedy);
int nbidx = 0;
for (int n = 0; n < NUM_NEIGHBORS; n++)
{
const int nbx = x + NEIGHBOR_DX[n];
const int nby = y + NEIGHBOR_DY[n];
if (nbx < 0 || nbx >= gridw || nby < 0 || nby >= gridh)
continue;
neighbors(i, nbidx++) = nby * gridw + nbx;
}
}
return nseeds;
}
static cv::Point2f intersection(const cv::Point2f& u1, const cv::Point2f& u2, const cv::Point2f& v1, const cv::Point2f& v2)
{
cv::Point2d ans = u1;
const double t = ((u1.x - v1.x) * (v1.y - v2.y) - (u1.y - v1.y) * (v1.x - v2.x)) /
((u1.x - u2.x) * (v1.y - v2.y) - (u1.y - u2.y) * (v1.x - v2.x));
ans.x += (u2.x - u1.x) * t;
ans.y += (u2.y - u1.y) * t;
return cv::Point2f(ans);
}
// circle center containing a triangular
static cv::Point2f circumcenter(const cv::Point2f& a, const cv::Point2f& b, const cv::Point2f& c)
{
cv::Point2f ua, ub, va, vb;
ua.x = (a.x + b.x) / 2;
ua.y = (a.y + b.y) / 2;
ub.x = ua.x - a.y + b.y;
ub.y = ua.y + a.x - b.x;
va.x = (a.x + c.x) / 2;
va.y = (a.y + c.y) / 2;
vb.x = va.x - a.y + c.y;
vb.y = va.y + a.x - c.x;
return intersection(ua, ub, va, vb);
}
static double dist(const cv::Point2f& p1, const cv::Point2f& p2)
{
return cv::norm(p2 - p1);
}
static float minimalCircle(const cv::Point2f p[], int n)
{
const double eps = 1e-6;
// center and radius of the circle
cv::Point2f o;
double r;
int i, j, k;
o = p[0];
r = 0;
for (i = 1; i < n; i++)
{
if (dist(p[i], o) - r > eps)
{
o = p[i];
r = 0;
for (j = 0; j < i; j++)
{
if (dist(p[j], o) - r > eps)
{
o.x = 0.5f * (p[i].x + p[j].x);
o.y = 0.5f * (p[i].y + p[j].y);
r = dist(o, p[j]);
for (k = 0; k < j; k++)
{
if (dist(o, p[k]) - r > eps)
{
o = circumcenter(p[i], p[j], p[k]);
r = dist(o, p[k]);
}
}
}
}
}
}
return static_cast<float>(r);
}
static void updateSearchRadius(const cv::Mat2f& flow, const cv::Mat1i& neighbors, cv::Mat1f& searchRadius)
{
const int nseeds = flow.rows;
cv::Point2f flows[NUM_NEIGHBORS + 1];
for (int i = 0; i < nseeds; i++)
{
flows[0] = flow(i);
int count = 1;
for (int n = 0; n < NUM_NEIGHBORS && neighbors(i, n) >= 0; n++)
flows[count++] = flow(neighbors(i, n));
searchRadius(i) = minimalCircle(flows, count);
}
}
struct MatchingCost
{
MatchingCost(const cv::Mat& desc1, const cv::Mat& desc2, const cv::Size& imgSize)
: desc1(desc1), desc2(desc2), w(imgSize.width), h(imgSize.height), ch(desc1.cols)
{
CV_Assert(desc1.type() == CV_8U && desc2.type() == CV_8U);
CV_Assert(desc1.size() == desc2.size());
#ifdef WITH_SSE
CV_Assert(ch * 255 < std::numeric_limits<ushort>::max()); // confirm maximum SAD does not exceed 16bit
#endif
}
inline int compute(const cv::Vec2f& pt, const cv::Vec2f& flow) const
{
const int x1 = clamp(static_cast<int>(pt[0] + 0.5f), 0, w - 1);
const int y1 = clamp(static_cast<int>(pt[1] + 0.5f), 0, h - 1);
const int x2 = clamp(static_cast<int>(pt[0] + flow[0] + 0.5f), 0, w - 1);
const int y2 = clamp(static_cast<int>(pt[1] + flow[1] + 0.5f), 0, h - 1);
const int idx1 = y1 * w + x1;
const int idx2 = y2 * w + x2;
const uchar* ptr1 = desc1.ptr<uchar>(idx1);
const uchar* ptr2 = desc2.ptr<uchar>(idx2);
#ifdef WITH_SSE
const __m128i* vptr1 = reinterpret_cast<const __m128i*>(ptr1);
const __m128i* vptr2 = reinterpret_cast<const __m128i*>(ptr2);
__m128i vsum = _mm_setzero_si128();
for (int i = 0; i < ch / 16; i++)
{
const __m128i v1 = _mm_loadu_si128(vptr1++);
const __m128i v2 = _mm_loadu_si128(vptr2++);
const __m128i vsad = _mm_sad_epu8(v1, v2);
vsum = _mm_adds_epu16(vsum, vsad);
}
int diff = _mm_extract_epi16(vsum, 0) + _mm_extract_epi16(vsum, 4);
for (int i = (ch & ~0x0f); i < ch; i++)
diff += std::abs(ptr1[i] - ptr2[i]);
#else
int diff = 0;
for (int i = 0; i < ch; i++)
diff += std::abs(ptr1[i] - ptr2[i]);
#endif
return diff;
}
const cv::Mat& desc1;
const cv::Mat& desc2;
int w, h, ch;
};
class RandomFlow
{
public:
RandomFlow(uint64 seed = 0) : rng_(seed) {}
inline cv::Vec2f generate(int radius)
{
const float x = static_cast<float>(rng_.uniform(-radius, radius + 1));
const float y = static_cast<float>(rng_.uniform(-radius, radius + 1));
return cv::Vec2f(x, y);
}
private:
cv::RNG rng_;
};
class Propagation
{
public:
Propagation(int maxIters = 8, double stopIterRatio = 0.05) : maxIters_(maxIters), stopIterRatio_(stopIterRatio) {}
void operator()(cv::Mat2f& flow, const cv::Mat& desc1, const cv::Mat& desc2, const cv::Size& imgSize,
const cv::Mat2f& seeds, const cv::Mat1i& neighbors, const cv::Mat1f& radius)
{
CV_Assert(desc1.type() == CV_8U && desc2.type() == CV_8U);
const double eps = 1e-6;
const int nseeds = seeds.rows;
RandomFlow rf;
MatchingCost C(desc1, desc2, imgSize);
std::vector<int> visited(nseeds), bestCosts(nseeds);
// init cost
for (int i = 0; i < nseeds; i++)
bestCosts[i] = C.compute(seeds(i), flow(i));
double lastUpdateRatio = 2;
for (int iter = 0; iter < maxIters_; iter++)
{
int updateCount = 0;
memset(visited.data(), 0, sizeof(int) * nseeds);
int i0 = 0, i1 = nseeds, step = 1;
if (iter % 2 == 1)
{
i0 = nseeds - 1; i1 = -1; step = -1;
}
for (int ic = i0; ic != i1; ic += step)
{
bool updated = false;
const cv::Vec2f pt = seeds(ic);
// Propagation: Improve current guess by trying instead correspondences from neighbors
for (int n = 0; n < NUM_NEIGHBORS && neighbors(ic, n) >= 0; n++)
{
const int in = neighbors(ic, n);
if (!visited[in])
continue;
const cv::Vec2f fc = flow(ic);
const cv::Vec2f ft = flow(in);
const cv::Vec2f df = ft - fc;
if (fabs(df[0]) < eps && fabs(df[1]) < eps)
continue;
const int cost = C.compute(pt, ft);
if (cost < bestCosts[ic])
{
bestCosts[ic] = cost;
flow(ic) = ft;
updated = true;
}
}
// Random search: Improve current guess by searching in boxes
// of exponentially decreasing size around the current best guess.
for (int mag = cvRound(radius(ic)); mag >= 1; mag /= 2)
{
/* Sampling window */
const cv::Vec2f fc = flow(ic);
const cv::Vec2f ft = fc + rf.generate(mag);
const cv::Vec2f df = ft - fc;
if (fabs(df[0]) < eps && fabs(df[1]) < eps)
continue;
const int cost = C.compute(pt, ft);
if (cost < bestCosts[ic])
{
bestCosts[ic] = cost;
flow(ic) = ft;
updated = true;
}
}
visited[ic] = 1;
if (updated)
updateCount++;
}
const double updateRatio = 1. * updateCount / nseeds;
if (updateRatio < stopIterRatio_ || lastUpdateRatio - updateRatio < 0.01)
break;
lastUpdateRatio = updateRatio;
}
}
private:
int maxIters_;
double stopIterRatio_;
};
class PyramidRandomSearch
{
public:
using Parameters = CoarseToFinePatchMatch::Parameters;
PyramidRandomSearch(const std::vector<cv::Mat>& I1s, const std::vector<cv::Mat>& I2s, const cv::Mat2f& seeds,
const cv::Mat1i& neighbors, const Parameters& param = Parameters()) : neighbors_(neighbors), param_(param)
{
init(I1s, I2s, seeds);
}
void init(const std::vector<cv::Mat>& I1s, const std::vector<cv::Mat>& I2s, const cv::Mat2f& seeds)
{
const int nseeds = seeds.rows;
const int nscales = static_cast<int>(I1s.size());
const std::vector<double> scales = makeScales(param_.scaleStep, nscales);
seeds_.assign(nscales, cv::Mat2f(nseeds, 1));
sizes_.assign(nscales, cv::Size());
maxRadius_.assign(nscales, 0);
for (int s = 0; s < nscales; s++)
{
const int w = I1s[s].cols;
const int h = I1s[s].rows;
seeds_[s] = cv::Mat2f(scales[s] * seeds);
sizes_[s] = cv::Size(w, h);
maxRadius_[s] = std::min(cvRound(param_.maxDisp * scales[s]), 32);
}
initRadius_ = cvRound(param_.maxDisp * scales[nscales - 1]);
}
cv::Mat2f compute(const std::vector<cv::Mat>& desc1, const std::vector<cv::Mat>& desc2)
{
const int nscales = static_cast<int>(desc1.size());
const int nseeds = seeds_[0].rows;
// random Initialization on coarsest level
RandomFlow rf;
cv::Mat2f flow(nseeds, 1);
for (int i = 0; i < nseeds; i++)
flow(i) = rf.generate(initRadius_);
// set the radius of coarsest level
cv::Mat1f radius(nseeds, 1);
radius = static_cast<float>(initRadius_);
Propagation propagate(param_.maxIters, param_.stopIterRatio);
// coarse-to-fine
for (int s = nscales - 1; s >= 0; s--)
{
propagate(flow, desc1[s], desc2[s], sizes_[s], seeds_[s], neighbors_, radius);
if (s > 0)
{
updateSearchRadius(flow, neighbors_, radius);
// scale the radius accordingly
radius = cv::max(1, cv::min(radius, maxRadius_[s]));
radius *= (1. / param_.scaleStep);
flow *= (1. / param_.scaleStep);
}
}
return flow;
}
private:
std::vector<cv::Mat2f> seeds_;
std::vector<cv::Size> sizes_;
std::vector<int> maxRadius_;
cv::Mat1i neighbors_;
int initRadius_;
Parameters param_;
};
static int crossCheck(const cv::Mat2f& seeds, const cv::Mat2f& flow1, const cv::Mat2f& flow2, std::vector<int>& valid,
int w, int h, int step, int maxDisp, int checkThreshold, int borderWidth)
{
const int nseeds = seeds.rows;
const int radius = step / 2;
const cv::Rect region(borderWidth, borderWidth, w - 2 * borderWidth, h - 2 * borderWidth);
cv::Mat1i labels(h, w);
for (int i = 0; i < nseeds; i++)
{
const int x0 = static_cast<int>(seeds(i)[0]);
const int y0 = static_cast<int>(seeds(i)[1]);
for (int dy = -radius; dy <= radius; dy++)
{
for (int dx = -radius; dx <= radius; dx++)
{
const int x = clamp(x0 + dx, 0, w - 1);
const int y = clamp(y0 + dy, 0, h - 1);
labels(y, x) = i;
}
}
}
valid.assign(nseeds, 1);
int nvalids = 0;
for (int i = 0; i < nseeds; i++)
{
const cv::Vec2f f1 = flow1(i);
const cv::Point pt1(seeds(i));
const cv::Point pt2(seeds(i) + f1);
if (!region.contains(pt1) || !region.contains(pt2) || cv::norm(f1) > maxDisp)
{
valid[i] = 0;
continue;
}
const cv::Vec2f f2 = flow2(labels(pt2));
if (cv::norm(f1 + f2) > checkThreshold)
{
valid[i] = 0;
continue;
}
nvalids++;
}
return nvalids;
}
CoarseToFinePatchMatch::CoarseToFinePatchMatch(const Parameters& param) : param_(param)
{
}
cv::Mat4f CoarseToFinePatchMatch::compute(const cv::Mat& I1, const cv::Mat& I2)
{
CV_Assert(I1.size() == I2.size() && I1.type() == CV_8U && I2.type() == CV_8U);
const int h = I1.rows;
const int w = I1.cols;
const int step = param_.step;
// construct pyramid
std::vector<cv::Mat> I1s, I2s;
constructPyramid(I1, I1s, param_.scaleStep, 30);
constructPyramid(I2, I2s, param_.scaleStep, 30);
const int nscales = static_cast<int>(I1s.size());
// get feature
auto daisy = cv::xfeatures2d::DAISY::create(5, 3, 4, 8, cv::xfeatures2d::DAISY::NRM_FULL, cv::noArray(), false, false);
std::vector<cv::Mat> desc1(nscales), desc2(nscales);
for (int s = 0; s < nscales; s++)
{
daisy->compute(I1s[s], desc1[s]);
daisy->compute(I2s[s], desc2[s]);
desc1[s].convertTo(desc1[s], CV_8U, 255);
desc2[s].convertTo(desc2[s], CV_8U, 255);
}
// make seeds and neighbors
cv::Mat2f seeds;
cv::Mat1i neighbors;
const int nseeds = makeSeedsAndNeighbors(w, h, step, seeds, neighbors);
// pyramid random search
PyramidRandomSearch search(I1s, I2s, seeds, neighbors, param_);
const cv::Mat2f flow1 = search.compute(desc1, desc2);
const cv::Mat2f flow2 = search.compute(desc2, desc1);
// cross check
std::vector<int> valid;
const int nmatches = crossCheck(seeds, flow1, flow2, valid, w, h, step, param_.maxDisp, param_.checkTh, param_.borderWidth);
// flow to match
cv::Mat4f matches(nmatches, 1);
int idx = 0;
for (int i = 0; i < nseeds; i++)
{
if (valid[i])
{
matches(idx)[0] = seeds(i)[0];
matches(idx)[1] = seeds(i)[1];
matches(idx)[2] = seeds(i)[0] + flow1(i)[0];
matches(idx)[3] = seeds(i)[1] + flow1(i)[1];
idx++;
}
}
return matches;
}