-
Notifications
You must be signed in to change notification settings - Fork 31
/
ARAPDataTerm.cc
213 lines (192 loc) · 7.6 KB
/
ARAPDataTerm.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
#include <ARAPDataTerm.h>
#include <vw/Image/Filter.h>
#include <vw/Image/ImageView.h>
#include <vw/Image/Transform.h>
#include <ceres/ceres.h>
using namespace vw;
double vw::stereo::gradient_cost_metric(ImageView<float> const& a,
ImageView<float> const& b,
float alpha, float t_col, float t_grad ) {
ImageView<float> agrad =
sqrt(square(derivative_filter(a, 1, 0)) + square(derivative_filter(a, 0, 1)));
ImageView<float> bgrad =
sqrt(square(derivative_filter(b, 1, 0)) + square(derivative_filter(b, 0, 1)));
double sum_of_error = 0;
float nalpha = 1.0 - alpha;
for (int j = 0; j < a.rows(); j++) {
for (int i = 0; i < a.cols(); i++) {
sum_of_error +=
nalpha * std::min(fabsf(a(i,j) - b(i,j)), t_col) +
alpha * std::min(fabsf(agrad(i,j) - bgrad(i,j)), t_grad);
}
}
return sum_of_error;
}
class DisparitySurfaceTransform : public TransformBase<DisparitySurfaceTransform> {
Vector<double, 10> const& surface;
Vector2 center;
public:
DisparitySurfaceTransform(Vector<double, 10> 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[0] +
surface[2] * dp[1] +
surface[3] * dp2[0] +
surface[4] * dp2[1],
surface[5] +
surface[6] * dp[0] +
surface[7] * dp[1] +
surface[8] * dp2[0] +
surface[9] * dp2[1]);
}
};
double vw::stereo::evaluate_superpixel(ImageView<float> const& a,
ImageView<float> const& b,
BBox2i const& a_superpixel,
Vector2 const& a_barycenter,
Vector<double, 10> const& surface) {
ImageView<float> a_crop = crop(a, a_superpixel);
ImageView<float> b_crop =
crop(transform(b, DisparitySurfaceTransform(surface,
a_barycenter)),
a_superpixel);
return stereo::gradient_cost_metric(a_crop, b_crop);
}
double vw::stereo::evaluate_intermediate_term(double theta,
ImageView<Vector2f> const& u,
BBox2i const& a_superpixel,
Vector2 const& a_barycenter,
Vector<double, 10> const& surface) {
if (theta < 1e-3) {
// Practically zero
return 0;
}
// Render a disparity
double sum_error = 0;
for (int j = 0; j < a_superpixel.height(); j++ ) {
for (int i = 0; i < a_superpixel.width(); i++ ) {
Vector2 dp = Vector2(i,j) + Vector2(a_superpixel.min()) - a_barycenter;
Vector2 dp2 = elem_prod(dp, dp);
Vector2 disp(
surface[0] +
surface[1] * dp[0] +
surface[2] * dp[1] +
surface[3] * dp2[0] +
surface[4] * dp2[1],
surface[5] +
surface[6] * dp[0] +
surface[7] * dp[1] +
surface[8] * dp2[0] +
surface[9] * dp2[1]);
sum_error += norm_2_sqr(disp - u(i + a_superpixel.min()[0],
j + a_superpixel.min()[1]));
}
}
return theta * sum_error;
}
struct QuadraticSurfaceFit {
QuadraticSurfaceFit(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(y_) +
polynomial_dx[3] * T(xx_) +
polynomial_dx[4] * T(yy_)
);
residuals[1] = T(obs_dy_) -
(polynomial_dy[0] +
polynomial_dy[1] * T(x_) +
polynomial_dy[2] * T(y_) +
polynomial_dy[3] * T(xx_) +
polynomial_dy[4] * T(yy_)
);
return true;
}
double obs_dx_, obs_dy_, x_, y_, xx_, yy_;
};
void vw::stereo::fit_surface_superpixel(ImageView<PixelMask<Vector2i> > const& a_disp,
BBox2i const& a_subpixel,
Vector2 const& a_barycenter,
Vector<double, 10> & 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<QuadraticSurfaceFit, 2, 5, 5>
(new QuadraticSurfaceFit
(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[5]);
}
}
}
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 vw::stereo::define_superpixels(ImageView<PixelMask<Vector2i> > const& a_disp,
std::vector<std::pair<BBox2i, Vector2> > & superpixels,
std::vector<Vector<double, 10> > & 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]);
}
}
void vw::stereo::render_disparity_image(std::vector<std::pair<BBox2i, Vector2> > const& superpixels,
std::vector<Vector<double, 10> > const& superpixel_surfaces,
ImageView<PixelMask<Vector2f> > & disp) {
BBox2i output_size;
for (std::vector<std::pair<BBox2i, Vector2> >::const_iterator it =
superpixels.begin();
it != superpixels.end(); it++ ) {
output_size.grow(it->first);
}
// Render an image of what the surfaces represent
disp.set_size(output_size.width(), output_size.height());
fill(disp, PixelMask<Vector2f>(Vector2f()));
for (size_t s = 0; s < superpixels.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);
disp(i, j)[0] =
superpixel_surfaces[s][0] +
superpixel_surfaces[s][1] * dp[0] +
superpixel_surfaces[s][2] * dp[1] +
superpixel_surfaces[s][3] * dp2[0] +
superpixel_surfaces[s][4] * dp2[1];
disp(i, j)[1] =
superpixel_surfaces[s][5] +
superpixel_surfaces[s][6] * dp[0] +
superpixel_surfaces[s][7] * dp[1] +
superpixel_surfaces[s][8] * dp2[0] +
superpixel_surfaces[s][9] * dp2[1];
}
}
}
}