forked from zmoratto/PatchMatch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
SurfaceFitWCostView.cc
290 lines (264 loc) · 10.2 KB
/
SurfaceFitWCostView.cc
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
#include <SurfaceFitWCostView.h>
#include <NelderMead.h>
#include <vw/Math/BBox.h>
#include <vw/Image/Transform.h>
#include <vw/Image/ImageMath.h>
#include <vw/Image/Statistics.h>
#include <vw/Stereo/DisparityMap.h>
#include <ceres/ceres.h>
using namespace vw;
struct Polynomial2DSurfaceFit {
Polynomial2DSurfaceFit(double obs_dx, double obs_dy,
double x, double y) :
obs_dx_(obs_dx), obs_dy_(obs_dy), x_(x), y_(y), xx_(x*x), yy_(y*y) {}
template <typename T>
bool operator()(const T* const polynomial_dx,
const T* const polynomial_dy,
T* residuals) const {
residuals[0] = T(obs_dx_) -
(polynomial_dx[0] +
polynomial_dx[1] * T(x_) +
polynomial_dx[2] * T(xx_) +
polynomial_dx[3] * T(y_) +
polynomial_dx[4] * T(y_) * T(x_) +
polynomial_dx[5] * T(y_) * T(xx_) +
polynomial_dx[6] * T(yy_) +
polynomial_dx[7] * T(yy_) * T(x_) +
polynomial_dx[8] * T(yy_) * T(xx_)
);
residuals[1] = T(obs_dy_) -
(polynomial_dy[0] +
polynomial_dy[1] * T(x_) +
polynomial_dy[2] * T(xx_) +
polynomial_dy[3] * T(y_) +
polynomial_dy[4] * T(y_) * T(x_) +
polynomial_dy[5] * T(y_) * T(xx_) +
polynomial_dy[6] * T(yy_) +
polynomial_dy[7] * T(yy_) * T(x_) +
polynomial_dy[8] * T(yy_) * T(xx_)
);
return true;
}
double obs_dx_, obs_dy_, x_, y_, xx_, yy_;
};
void fit_surface_superpixel(ImageView<PixelMask<Vector2i> > const& a_disp,
BBox2i const& a_subpixel,
Vector2 const& a_barycenter,
Vector<double, 18> & surface) {
ceres::Problem problem;
for (int j = a_subpixel.min()[1]; j < a_subpixel.max()[1]; j++) {
for (int i = a_subpixel.min()[0]; i < a_subpixel.max()[0]; i++) {
if (is_valid(a_disp(i,j))) {
problem.AddResidualBlock
(new ceres::AutoDiffCostFunction<Polynomial2DSurfaceFit, 2, 9, 9>
(new Polynomial2DSurfaceFit
(a_disp(i,j).child()[0], a_disp(i,j).child()[1],
double(i) - a_barycenter[0],
double(j) - a_barycenter[1])),
new ceres::CauchyLoss(3),
&surface[0], &surface[9]);
}
}
}
ceres::Solver::Options options;
options.max_num_iterations = 300;
options.minimizer_progress_to_stdout = false;
ceres::Solver::Summary summary;
ceres::Solve(options, &problem, &summary);
if (summary.termination_type == ceres::NO_CONVERGENCE) {
std::fill(surface.begin(), surface.end(), 0);
}
}
void define_superpixels(ImageView<PixelMask<Vector2i> > const& a_disp,
std::vector<std::pair<BBox2i, Vector2> > & superpixels,
std::vector<Vector<double, 18> > & superpixel_surfaces) {
superpixel_surfaces.resize(superpixels.size());
for (size_t i = 0; i < superpixels.size(); i++ ) {
fit_surface_superpixel(a_disp,
superpixels[i].first,
superpixels[i].second,
superpixel_surfaces[i]);
}
}
class DisparityQuadSurfaceTransform : public TransformBase<DisparityQuadSurfaceTransform> {
Vector<double, 18> const& surface;
Vector2 center;
public:
DisparityQuadSurfaceTransform(Vector<double, 18> const& s,
Vector2 const& c) :
surface(s), center(c) {}
inline Vector2 reverse(const Vector2 &p ) const {
// Give a destination pixel ... return the pixel that should be the source
Vector2 dp = p - center;
Vector2 dp2 = elem_prod(dp, dp);
return p +
Vector2(surface[0] +
surface[1] * dp.x() +
surface[2] * dp2.x() +
surface[3] * dp.y() +
surface[4] * dp.y() * dp.x() +
surface[5] * dp.y() * dp2.x() +
surface[6] * dp2.y() +
surface[7] * dp2.y() * dp.x() +
surface[8] * dp2.y() * dp2.x(),
surface[9] +
surface[10] * dp.x() +
surface[11] * dp2.x() +
surface[12] * dp.y() +
surface[13] * dp.y() * dp.x() +
surface[14] * dp.y() * dp2.x() +
surface[15] * dp2.y() +
surface[16] * dp2.y() * dp.x() +
surface[17] * dp2.y() * dp2.x());
}
};
struct NCCQuadraticFunctor {
ImageView<float> const& left, right;
std::pair<BBox2i, Vector2> const& superpixel;
NCCQuadraticFunctor(ImageView<float> const& a,
ImageView<float> const& b,
std::pair<BBox2i, Vector2> const& s ) :
left(a), right(b), superpixel(s) {}
double operator()(Vector<double, 18> const& surface) const {
ImageView<float> left_kernel = crop(left, superpixel.first);
ImageView<float> right_kernel =
crop(transform(right, DisparityQuadSurfaceTransform(surface, superpixel.second)),
superpixel.first);
float cov_lr =
sum_of_pixel_values
(left_kernel * right_kernel);
float cov_ll =
sum_of_pixel_values
(left_kernel * left_kernel);
float cov_rr =
sum_of_pixel_values
(right_kernel * right_kernel);
// NCC Cost function here
return 1 - cov_lr / sqrt(cov_ll * cov_rr);
}
};
void
vw::stereo::SurfaceFitWCost(ImageView<PixelMask<Vector2f> > surface,
ImageView<float> left, ImageView<float> right) {
// Define our super pixels
std::vector<BBox2i> box_vec =
image_blocks(surface, 64, 64);
std::vector<std::pair<BBox2i, Vector2> > superpixels;
superpixels.reserve(box_vec.size());
for (std::vector<BBox2i>::iterator it =
box_vec.begin(); it != box_vec.end(); it++) {
superpixels.push_back
(std::make_pair
(*it,
Vector2(it->min()) +
Vector2(it->size()) / 2));
}
std::cout << "Number of superpixels: "
<< superpixels.size() << std::endl;
std::vector<Vector<double, 18> > superpixel_surfaces;
define_superpixels(surface,
superpixels,
superpixel_surfaces);
// Look through the fitted surfaces and identify ones that seem to
// be obvious outliers.
BBox2i disp_range = stereo::get_disparity_range(surface);
for (int s = 0; s < superpixels.size(); s++ ) {
if (!disp_range.contains(Vector2i(superpixel_surfaces[s][0],
superpixel_surfaces[s][9]))) {
std::cout << "Zeroing " << s << " " << superpixel_surfaces[s] << std::endl;
std::fill(superpixel_surfaces[s].begin(),
superpixel_surfaces[s].end(), 0);
}
}
// Iterate through the fitted surfaces and refine them so that they reduce a general NCC cost
int width = surface.cols() / 64;
for (int s = 0; s < superpixels.size(); s++ ) {
std::cout << s << std::endl;
NCCQuadraticFunctor functor(left, right,
superpixels[s]);
std::cout << "starting cost: " << functor(superpixel_surfaces[s]) << std::endl;
std::cout << superpixel_surfaces[s] << std::endl;
Vector<double, 18> seeds[19];
int write_index = 0;
seeds[write_index++] = superpixel_surfaces[s];
if (s > 0) {
seeds[write_index++] = superpixel_surfaces[s-1];
}
if (s < superpixel_surfaces.size() - 1) {
seeds[write_index++] = superpixel_surfaces[s+1];
}
if (s >= width + 1) {
seeds[write_index++] = superpixel_surfaces[s - width - 1];
}
if (s >= width) {
seeds[write_index++] = superpixel_surfaces[s - width];
}
if (s >= width - 1) {
seeds[write_index++] = superpixel_surfaces[s - width + 1];
}
if (s < superpixel_surfaces.size() - width) {
seeds[write_index++] = superpixel_surfaces[s + width - 1];
}
if (s < superpixel_surfaces.size() - width - 1) {
seeds[write_index++] = superpixel_surfaces[s + width];
}
if (s < superpixel_surfaces.size() - width - 2) {
seeds[write_index++] = superpixel_surfaces[s + width + 1];
}
if (s > 1) {
seeds[write_index++] = superpixel_surfaces[s-2];
}
if (s < superpixel_surfaces.size() - 2) {
seeds[write_index++] = superpixel_surfaces[s+2];
}
if (s >= 2 * width) {
seeds[write_index++] = superpixel_surfaces[s - 2 * width];
}
if (s < superpixel_surfaces.size() - 2 * width - 1) {
seeds[write_index++] = superpixel_surfaces[s + 2 * width];
}
// Minimium, this fills in 5 elements .. worse case it fills in 14 elements
while (write_index < 19) {
seeds[write_index] = superpixel_surfaces[s];
seeds[write_index][18 - write_index] += 0.1;
write_index++;
}
stereo::Amoeba<18> amoeba(1e-4);
superpixel_surfaces[s] =
amoeba.minimize(seeds, functor);
std::cout << "ending cost: " << functor(superpixel_surfaces[s]) << std::endl;
std::cout << superpixel_surfaces[s] << std::endl;
}
// Render back out to our input so that it has our surface fit
fill(surface, PixelMask<Vector2f>(Vector2f()));
for (size_t s = 0; s < superpixel_surfaces.size(); s++ ) {
for (int j = superpixels[s].first.min()[1];
j < superpixels[s].first.max()[1]; j++ ) {
for (int i = superpixels[s].first.min()[0];
i < superpixels[s].first.max()[0]; i++ ) {
Vector2 dp = Vector2(i,j) - superpixels[s].second;
Vector2 dp2 = elem_prod(dp, dp);
surface(i, j)[0] =
superpixel_surfaces[s][0] +
superpixel_surfaces[s][1] * dp.x() +
superpixel_surfaces[s][2] * dp2.x() +
superpixel_surfaces[s][3] * dp.y() +
superpixel_surfaces[s][4] * dp.y() * dp.x() +
superpixel_surfaces[s][5] * dp.y() * dp2.x() +
superpixel_surfaces[s][6] * dp2.y() +
superpixel_surfaces[s][7] * dp2.y() * dp.x() +
superpixel_surfaces[s][8] * dp2.y() * dp2.x();
surface(i, j)[1] =
superpixel_surfaces[s][9] +
superpixel_surfaces[s][10] * dp.x() +
superpixel_surfaces[s][11] * dp2.x() +
superpixel_surfaces[s][12] * dp.y() +
superpixel_surfaces[s][13] * dp.y() * dp.x() +
superpixel_surfaces[s][14] * dp.y() * dp2.x() +
superpixel_surfaces[s][15] * dp2.y() +
superpixel_surfaces[s][16] * dp2.y() * dp.x() +
superpixel_surfaces[s][17] * dp2.y() * dp2.x();
}
}
}
}