Merge "Adds code for corner detection and ransac" into nextgen

This commit is contained in:
Deb Mukherjee 2015-02-10 08:18:03 -08:00 committed by Gerrit Code Review
commit 2aef964519
9 changed files with 4696 additions and 0 deletions

File diff suppressed because it is too large Load Diff

View File

@ -0,0 +1,24 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*
*/
#ifndef VP9_ENCODER_VP9_CORNER_DETECT_H
#define VP9_ENCODER_VP9_CORNER_DETECT_H
#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
int HarrisCornerDetect(unsigned char *buf, int width, int height, int stride,
int *points, int max_points);
int FastCornerDetect(unsigned char *buf, int width, int height, int stride,
int *points, int max_points);
#endif // VP9_ENCODER_VP9_CORNER_DETECT_H

View File

@ -0,0 +1,224 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*
*/
#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
#include <math.h>
#include "vp9_corner_match.h"
#define MATCH_SZ 21
#define MATCH_SZ_BY2 ((MATCH_SZ - 1)/2)
#define MATCH_SZ_SQ (MATCH_SZ * MATCH_SZ)
#define SEARCH_SZ 9
#define SEARCH_SZ_BY2 ((SEARCH_SZ - 1)/2)
#define THRESHOLD_NCC 0.80
typedef struct {
int x, y;
int rx, ry;
} correspondence;
static double compute_variance(unsigned char *im, int stride,
int x, int y, double *mean) {
double sum = 0.0;
double sumsq = 0.0;
double var;
int i, j;
for (i = 0; i < MATCH_SZ; ++i)
for (j = 0; j < MATCH_SZ; ++j) {
sum += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
sumsq += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)] *
im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
}
var = (sumsq * MATCH_SZ_SQ - sum * sum) / (MATCH_SZ_SQ * MATCH_SZ_SQ);
if (mean) *mean = sum / MATCH_SZ_SQ;
return var;
}
static double compute_cross_correlation(unsigned char *im1, int stride1,
int x1, int y1,
unsigned char *im2, int stride2,
int x2, int y2) {
double sum1 = 0;
double sum2 = 0;
double cross = 0;
double corr;
int i, j;
for (i = 0; i < MATCH_SZ; ++i)
for (j = 0; j < MATCH_SZ; ++j) {
sum1 += im1[(i + y1 - MATCH_SZ_BY2) * stride1 + (j + x1 - MATCH_SZ_BY2)];
sum2 += im2[(i + y2 - MATCH_SZ_BY2) * stride2 + (j + x2 - MATCH_SZ_BY2)];
cross += im1[(i + y1 - MATCH_SZ_BY2) * stride1 + (j + x1 - MATCH_SZ_BY2)] *
im2[(i + y2 - MATCH_SZ_BY2) * stride2 + (j + x2 - MATCH_SZ_BY2)];
}
corr = (cross * MATCH_SZ_SQ - sum1 * sum2) / (MATCH_SZ_SQ * MATCH_SZ_SQ);
return corr;
}
static int is_eligible_point(double pointx, double pointy,
int width, int height) {
return (pointx >= MATCH_SZ_BY2 && pointy >= MATCH_SZ_BY2 &&
pointx + MATCH_SZ_BY2 < width && pointy + MATCH_SZ_BY2 < height);
}
static int is_eligible_distance(double point1x, double point1y,
double point2x, double point2y,
int width, int height) {
const int thresh = (width < height ? height : width) >> 4;
return ((point1x - point2x) * (point1x - point2x) +
(point1y - point2y) * (point1y - point2y)) <= thresh * thresh;
}
static void improve_correspondence(unsigned char *frm, unsigned char *ref,
int width, int height,
int frm_stride, int ref_stride,
correspondence *correspondences,
int num_correspondences) {
int i;
for (i = 0; i < num_correspondences; ++i) {
double template_norm = compute_variance(frm, frm_stride,
correspondences[i].x,
correspondences[i].y, NULL);
int x, y, best_x = 0, best_y = 0;
double best_match_ncc = 0.0;
for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y) {
for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
double match_ncc;
double subimage_norm;
if (!is_eligible_point(correspondences[i].rx + x,
correspondences[i].ry + y,
width, height))
continue;
if (!is_eligible_distance(
correspondences[i].x, correspondences[i].y,
correspondences[i].rx + x, correspondences[i].ry + y,
width, height))
continue;
subimage_norm = compute_variance(ref, ref_stride,
correspondences[i].rx + x,
correspondences[i].ry + y, NULL);
match_ncc = compute_cross_correlation(
frm, frm_stride,
correspondences[i].x, correspondences[i].y,
ref, ref_stride,
correspondences[i].rx + x, correspondences[i].ry + y) /
sqrt(template_norm * subimage_norm);
if (match_ncc > best_match_ncc) {
best_match_ncc = match_ncc;
best_y = y;
best_x = x;
}
}
}
correspondences[i].rx += best_x;
correspondences[i].ry += best_y;
}
for (i = 0; i < num_correspondences; ++i) {
double template_norm = compute_variance(
ref, ref_stride, correspondences[i].rx, correspondences[i].ry, NULL);
int x, y, best_x = 0, best_y = 0;
double best_match_ncc = 0.0;
for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y)
for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
double match_ncc;
double subimage_norm;
if (!is_eligible_point(correspondences[i].x + x,
correspondences[i].y + y, width, height))
continue;
if (!is_eligible_distance(
correspondences[i].x + x, correspondences[i].y + y,
correspondences[i].rx, correspondences[i].ry, width, height))
continue;
subimage_norm = compute_variance(
frm, frm_stride,
correspondences[i].x + x, correspondences[i].y + y, NULL);
match_ncc = compute_cross_correlation(
frm, frm_stride, correspondences[i].x + x, correspondences[i].y + y,
ref, ref_stride, correspondences[i].rx, correspondences[i].ry) /
sqrt(template_norm * subimage_norm);
if (match_ncc > best_match_ncc) {
best_match_ncc = match_ncc;
best_y = y;
best_x = x;
}
}
correspondences[i].x += best_x;
correspondences[i].y += best_y;
}
}
int determine_correspondence(unsigned char *frm,
int *frm_corners, int num_frm_corners,
unsigned char *ref,
int *ref_corners, int num_ref_corners,
int width, int height,
int frm_stride, int ref_stride,
int *correspondence_pts) {
// Debargha: Improve this to include 2-way match
int i, j;
correspondence *correspondences = (correspondence *)correspondence_pts;
int num_correspondences = 0;
for (i = 0; i < num_frm_corners; ++i) {
double best_match_ncc = 0.0;
double template_norm;
int best_match_j = -1;
if (!is_eligible_point(frm_corners[2 * i], frm_corners[2 * i + 1],
width, height))
continue;
template_norm = compute_variance(
frm, frm_stride, frm_corners[2 * i], frm_corners[2*i+1], NULL);
for (j = 0; j < num_ref_corners; ++j) {
double match_ncc;
double subimage_norm;
if (!is_eligible_point(ref_corners[2 * j], ref_corners[2 * j + 1],
width, height))
continue;
if (!is_eligible_distance(frm_corners[2 * i], frm_corners[2 * i + 1],
ref_corners[2 * j], ref_corners[2 * j + 1],
width, height))
continue;
subimage_norm = compute_variance(
ref, ref_stride, ref_corners[2*j], ref_corners[2 * j + 1], NULL);
match_ncc = compute_cross_correlation(frm, frm_stride,
frm_corners[2 * i],
frm_corners[2 * i + 1],
ref, ref_stride,
ref_corners[2 * j],
ref_corners[2 * j + 1]) /
sqrt(template_norm * subimage_norm);
if (match_ncc > best_match_ncc) {
best_match_ncc = match_ncc;
best_match_j = j;
}
}
if (best_match_ncc > THRESHOLD_NCC) {
correspondences[num_correspondences].x = frm_corners[2 * i];
correspondences[num_correspondences].y = frm_corners[2 * i + 1];
correspondences[num_correspondences].rx = ref_corners[2 * best_match_j];
correspondences[num_correspondences].ry =
ref_corners[2 * best_match_j + 1];
/*
printf(" %d %d %d %d\n",
correspondences[num_correspondences].x,
correspondences[num_correspondences].y,
correspondences[num_correspondences].rx,
correspondences[num_correspondences].ry);
*/
num_correspondences++;
}
}
improve_correspondence(frm, ref, width, height, frm_stride, ref_stride,
correspondences, num_correspondences);
return num_correspondences;
}

View File

@ -0,0 +1,27 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*
*/
#ifndef VP9_ENCODER_VP9_CORNER_MATCH_H
#define VP9_ENCODER_VP9_CORNER_MATCH_H
#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
int determine_correspondence(unsigned char *frm,
int *frm_corners, int num_frm_corners,
unsigned char *ref,
int *ref_corners, int num_ref_corners,
int width, int height,
int frm_stride, int ref_stride,
int *correspondence_pts);
#endif // VP9_ENCODER_VP9_CORNER_MATCH_H

View File

@ -0,0 +1,339 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
#include <math.h>
#include <assert.h>
#include "vp9_corner_detect.h"
#include "vp9_corner_match.h"
#include "vp9_ransac.h"
#include "vp9_global_motion.h"
// #define VERBOSE
// Default is Harris
#define USE_FAST_CORNER
#define MIN_INLIER_PROB 0.1
#define MAX_CORNERS 4096
inline int get_numparams(TransformationType type) {
switch (type) {
case HOMOGRAPHY:
return 9;
case AFFINE:
return 6;
case ROTZOOM:
return 4;
default:
assert(0);
return 0;
}
}
inline ransacType get_ransacType(TransformationType type) {
switch (type) {
case HOMOGRAPHY:
return ransacHomography;
case AFFINE:
return ransacAffine;
case ROTZOOM:
return ransacRotZoom;
default:
assert(0);
return NULL;
}
}
inline projectPointsType get_projectPointsType(TransformationType type) {
switch (type) {
case HOMOGRAPHY:
return projectPointsHomography;
case AFFINE:
return projectPointsAffine;
case ROTZOOM:
return projectPointsRotZoom;
default:
assert(0);
return NULL;
}
}
static double compute_error_score(TransformationType type,
int *points1, int stride1,
int *points2, int stride2,
int npoints, double *H,
int *map) {
int i, n = 0;
double ot[2], pt[2];
int *mp1 = points1;
int *mp2 = points2;
double sqerr = 0.0;
const int numparams = get_numparams(type);
projectPointsType projectPoints = get_projectPointsType(type);
if (projectPoints == NULL) return -1.0;
if (map) {
for (i = 0; i < npoints; ++i, mp1+=stride1, mp2+=stride2) {
if (map[i] != -1) {
ot[0] = mp1[0];
ot[1] = mp1[1];
projectPoints(&H[map[i] * numparams], ot, pt, 1, stride1, stride2);
sqerr += (pt[0] - mp2[0]) * (pt[0] - mp2[0]) +
(pt[1] - mp2[1]) * (pt[1] - mp2[1]);
n++;
}
}
} else {
for (i = 0; i < npoints; ++i, mp1+=stride1, mp2+=stride2) {
ot[0] = mp1[0];
ot[1] = mp1[1];
projectPoints(H, ot, pt, 1, stride1, stride2);
sqerr += (pt[0] - mp2[0]) * (pt[0] - mp2[0]) +
(pt[1] - mp2[1]) * (pt[1] - mp2[1]);
n++;
}
}
return sqrt(sqerr / n);
}
static int compute_global_motion_single(TransformationType type,
int *correspondences,
int num_correspondences,
double *H,
int *inlier_map) {
double *mp, *matched_points;
int *cp = correspondences;
int i, result;
int num_inliers = 0;
ransacType ransac = get_ransacType(type);
if (ransac == NULL)
return 0;
matched_points =
(double *)malloc(4 * num_correspondences * sizeof(double));
for (mp = matched_points, cp = correspondences, i = 0;
i < num_correspondences; ++i) {
*mp++ = *cp++;
*mp++ = *cp++;
*mp++ = *cp++;
*mp++ = *cp++;
}
result = ransac(matched_points, num_correspondences,
&num_inliers, inlier_map, H);
if (!result && num_inliers < MIN_INLIER_PROB * num_correspondences) {
result = 1;
num_inliers = 0;
}
if (!result) {
for (i = 0; i < num_correspondences; ++i) {
inlier_map[i] = inlier_map[i] - 1;
}
}
free(matched_points);
return num_inliers;
}
// Returns number of models actually returned: 1 - if success, 0 - if failure
int vp9_compute_global_motion_single_feature_based(TransformationType type,
unsigned char *frmbuf,
unsigned char *refbuf,
int width,
int height,
int frm_stride,
int ref_stride,
double *H) {
int num_frm_corners, num_ref_corners;
int num_correspondences;
int *correspondences;
int num_inliers;
int *inlier_map = NULL;
int frm_corners[2 * MAX_CORNERS], ref_corners[2 * MAX_CORNERS];
#ifdef USE_FAST_CORNER
num_frm_corners = FastCornerDetect(frmbuf, width, height, frm_stride,
frm_corners, MAX_CORNERS);
num_ref_corners = FastCornerDetect(refbuf, width, height, ref_stride,
ref_corners, MAX_CORNERS);
#else
num_frm_corners = HarrisCornerDetect(frmbuf, width, height, frm_stride,
frm_corners, MAX_CORNERS);
num_ref_corners = HarrisCornerDetect(refbuf, width, height, ref_stride,
ref_corners, MAX_CORNERS);
#endif
#ifdef VERBOSE
printf("Frame corners = %d\n", num_frm_corners);
printf("Reference corners = %d\n", num_ref_corners);
#endif
correspondences = (int *)malloc(
num_frm_corners * 4 * sizeof(int));
num_correspondences = determine_correspondence(frmbuf, (int*)frm_corners,
num_frm_corners,
refbuf, (int*)ref_corners,
num_ref_corners,
width, height,
frm_stride, ref_stride,
correspondences);
#ifdef VERBOSE
printf("Number of correspondences = %d\n", num_correspondences);
#endif
inlier_map = (int *)malloc(num_correspondences * sizeof(int));
num_inliers = compute_global_motion_single(type, correspondences,
num_correspondences, H,
inlier_map);
#ifdef VERBOSE
printf("Inliers = %d\n", num_inliers);
printf("Error Score (inliers) = %g\n",
compute_error_score(type, correspondences, 4, correspondences + 2, 4,
num_correspondences, H, inlier_map));
#endif
free(correspondences);
free(inlier_map);
return (num_inliers > 0);
}
static int compute_global_motion_multiple(TransformationType type,
int *correspondences,
int num_correspondences,
double *H,
int max_models,
double inlier_prob,
int *num_models,
int *processed_mask) {
int *cp = correspondences;
double *mp, *matched_points;
int *best_inlier_mask;
int i, result;
int num_points = 0;
int num_inliers = 0;
int num_inliers_sum = 0;
const int numparams = get_numparams(type);
ransacType ransac = get_ransacType(type);
if (ransac == NULL)
return 0;
matched_points =
(double *)malloc(4 * num_correspondences * sizeof(double));
best_inlier_mask =
(int *)malloc(num_correspondences * sizeof(int));
for (i = 0; i < num_correspondences; ++i)
processed_mask[i] = -1;
*num_models = 0;
while ((double)num_inliers_sum / (double)num_correspondences < inlier_prob &&
*num_models < max_models) {
num_points = 0;
for (mp = matched_points, cp = correspondences, i = 0;
i < num_correspondences; ++i) {
if (processed_mask[i] == -1) {
*mp++ = *cp++;
*mp++ = *cp++;
*mp++ = *cp++;
*mp++ = *cp++;
num_points++;
} else {
cp += 4;
}
}
result = ransac(matched_points, num_points,
&num_inliers, best_inlier_mask,
&H[(*num_models) * numparams]);
if (!result && num_inliers < MIN_INLIER_PROB * num_correspondences) {
result = 1;
num_inliers = 0;
}
if (!result) {
num_points = 0;
for (i = 0; i < num_correspondences; ++i) {
if (processed_mask[i] == -1) {
if (best_inlier_mask[num_points]) processed_mask[i] = *num_models;
num_points++;
}
}
num_inliers_sum += num_inliers;
(*num_models)++;
} else {
break;
}
}
free(matched_points);
free(best_inlier_mask);
return num_inliers_sum;
}
// Returns number of models actually returned
int vp9_compute_global_motion_multiple_feature_based(TransformationType type,
unsigned char *frmbuf,
unsigned char *refbuf,
int width,
int height,
int frm_stride,
int ref_stride,
int max_models,
double inlier_prob,
double *H) {
int num_frm_corners, num_ref_corners;
int num_correspondences;
int *correspondences;
int num_inliers;
int frm_corners[2 * MAX_CORNERS], ref_corners[2 * MAX_CORNERS];
int num_models = 0;
int *inlier_map = NULL;
#ifdef USE_FAST_CORNER
num_frm_corners = FastCornerDetect(frmbuf, width, height, frm_stride,
frm_corners, MAX_CORNERS);
num_ref_corners = FastCornerDetect(refbuf, width, height, ref_stride,
ref_corners, MAX_CORNERS);
#else
num_frm_corners = HarrisCornerDetect(frmbuf, width, height, frm_stride,
frm_corners, MAX_CORNERS);
num_ref_corners = HarrisCornerDetect(refbuf, width, height, ref_stride,
ref_corners, MAX_CORNERS);
#endif
#ifdef VERBOSE
printf("Frame corners = %d\n", num_frm_corners);
printf("Reference corners = %d\n", num_ref_corners);
#endif
correspondences = (int *)malloc(num_frm_corners * 4 * sizeof(int));
num_correspondences = determine_correspondence(frmbuf, (int*)frm_corners,
num_frm_corners,
refbuf, (int*)ref_corners,
num_ref_corners,
width, height,
frm_stride, ref_stride,
correspondences);
#ifdef VERBOSE
printf("Number of correspondences = %d\n", num_correspondences);
#endif
inlier_map = (int *)malloc(num_correspondences * sizeof(int));
num_inliers = compute_global_motion_multiple(type, correspondences,
num_correspondences, H,
max_models, inlier_prob,
&num_models, inlier_map);
#ifdef VERBOSE
printf("Models = %d, Inliers = %d\n", num_models, num_inliers);
if (num_models)
printf("Error Score (inliers) = %g\n",
compute_error_score(type, correspondences, 4, correspondences + 2, 4,
num_correspondences, H, inlier_map));
#endif
(void) num_inliers;
free(correspondences);
free(inlier_map);
return num_models;
}

View File

@ -0,0 +1,69 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef VP9_ENCODER_VP9_GLOBAL_MOTION_H
#define VP9_ENCODER_VP9_GLOBAL_MOTION_H
#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
#include <math.h>
#include <assert.h>
#include "vp9_corner_detect.h"
#include "vp9_corner_match.h"
#include "vp9_ransac.h"
// Default is Harris
#define USE_FAST_CORNER
#define MAX_CORNERS 4096
typedef enum {
UNKNOWN = -1,
HOMOGRAPHY, // homography, 8-parameter
AFFINE, // affine, 6-parameter
ROTZOOM // simplified affine with rotation and zoom only, 4-parameter
} TransformationType;
inline int get_numparams(TransformationType type);
inline ransacType get_ransacType(TransformationType type);
inline projectPointsType get_projectPointsType(TransformationType type);
// Returns number of models actually returned: 1 - if success, 0 - if failure
int vp9_compute_global_motion_single_feature_based(TransformationType type,
unsigned char *frm,
unsigned char *ref,
int width,
int height,
int frm_stride,
int ref_stride,
double *H);
// Returns number of models actually returned: 1+ - #models, 0 - if failure
// max_models is the maximum number of models returned
// inlier_prob is the probability of being inlier over all the models
// combined, beyond which no more models are computed.
// Ex. if max_models = 4, and inlier_prob = 0.8, and during the
// process three models together already cover more than 80% of the
// matching points, then only three models are returned.
int vp9_compute_global_motion_multiple_feature_based(TransformationType type,
unsigned char *frm,
unsigned char *ref,
int width,
int height,
int frm_stride,
int ref_stride,
int max_models,
double inlier_prob,
double *H);
#endif // VP9_ENCODER_VP9_GLOBAL_MOTION_H

951
vp9/encoder/vp9_ransac.c Normal file
View File

@ -0,0 +1,951 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <memory.h>
#include <math.h>
#include <time.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include "vp9_ransac.h"
#define MAX_PARAMDIM 9
#define MAX_MINPTS 4
// svdcmp
// Adopted from Numerical Recipes in C
static const double TINY_NEAR_ZERO = 1.0E-12;
static inline double SIGN(double a, double b) {
return ((b) >= 0 ? fabs(a) : -fabs(a));
}
static inline double PYTHAG(double a, double b) {
double absa, absb, ct;
absa = fabs(a);
absb = fabs(b);
if(absa > absb) {
ct = absb/absa;
return absa * sqrt(1.0 + ct*ct);
} else {
ct = absa/absb;
return (absb == 0) ? 0 : absb * sqrt(1.0 + ct*ct);
}
}
inline int IMIN(int a, int b) {
return (((a) < (b)) ? (a) : (b));
}
inline int IMAX(int a, int b) {
return (((a) < (b)) ? (b) : (a));
}
static void MultiplyMat(double *m1, double *m2, double *res,
const int M1, const int N1, const int N2) {
int timesInner = N1;
int timesRows = M1;
int timesCols = N2;
double sum;
int row, col, inner;
for( row = 0; row < timesRows; ++row ) {
for( col = 0; col < timesCols; ++col ) {
sum = 0;
for (inner = 0; inner < timesInner; ++inner )
sum += m1[row*N1 + inner] * m2[inner*N2 + col];
*(res++) = sum;
}
}
}
static int svdcmp_(double **u, int m, int n, double w[], double **v) {
const int max_its = 30;
int flag, i, its, j, jj, k, l, nm;
double anorm, c, f, g, h, s, scale, x, y, z;
double *rv1 = (double *)malloc(sizeof(double) * (n + 1));
g = scale = anorm = 0.0;
for (i=0;i<n;i++) {
l=i+1;
rv1[i]=scale*g;
g=s=scale=0.0;
if (i < m) {
for (k=i;k<m;k++) scale += fabs(u[k][i]);
if (scale) {
for (k=i; k<m; k++) {
u[k][i] /= scale;
s += u[k][i]*u[k][i];
}
f = u[i][i];
g = -SIGN(sqrt(s), f);
h=f*g-s;
u[i][i]=f-g;
for (j=l;j<n;j++) {
for (s=0.0,k=i;k<m;k++) s += u[k][i]*u[k][j];
f=s/h;
for (k=i;k<m;k++) u[k][j] += f*u[k][i];
}
for (k=i;k<m;k++) u[k][i] *= scale;
}
}
w[i]=scale *g;
g=s=scale=0.0;
if (i < m && i != n - 1) {
for (k=l;k<n;k++) scale += fabs(u[i][k]);
if (scale) {
for (k=l;k<n;k++) {
u[i][k] /= scale;
s += u[i][k]*u[i][k];
}
f=u[i][l];
g = -SIGN(sqrt(s),f);
h=f*g-s;
u[i][l]=f-g;
for (k=l;k<n;k++) rv1[k]=u[i][k]/h;
for (j=l;j<m;j++) {
for (s=0.0,k=l;k<n;k++) s += u[j][k]*u[i][k];
for (k=l;k<n;k++) u[j][k] += s*rv1[k];
}
for (k=l;k<n;k++) u[i][k] *= scale;
}
}
anorm=fmax(anorm,(fabs(w[i])+fabs(rv1[i])));
}
for (i=n-1;i>=0;i--) {
if (i < n - 1) {
if (g) {
for (j=l;j<n;j++) v[j][i]=(u[i][j]/u[i][l])/g;
for (j=l;j<n;j++) {
for (s=0.0,k=l;k<n;k++) s += u[i][k]*v[k][j];
for (k=l;k<n;k++) v[k][j] += s*v[k][i];
}
}
for (j=l;j<n;j++) v[i][j]=v[j][i]=0.0;
}
v[i][i]=1.0;
g=rv1[i];
l=i;
}
for (i = IMIN(m,n) - 1; i>=0; i--) {
l=i+1;
g=w[i];
for (j=l;j<n;j++) u[i][j]=0.0;
if (g) {
g=1.0/g;
for (j=l;j<n;j++) {
for (s=0.0,k=l;k<m;k++) s += u[k][i]*u[k][j];
f = (s/u[i][i])*g;
for (k=i;k<m;k++) u[k][j] += f*u[k][i];
}
for (j=i;j<m;j++) u[j][i] *= g;
} else for (j=i;j<m;j++) u[j][i]=0.0;
++u[i][i];
}
for (k=n-1;k>=0;k--) {
for (its=0;its<max_its;its++) {
flag=1;
for (l=k;l>=0;l--) {
nm=l-1;
if ((double)(fabs(rv1[l])+anorm) == anorm || nm < 0) {
flag=0;
break;
}
if ((double)(fabs(w[nm])+anorm) == anorm) break;
}
if (flag) {
c=0.0;
s=1.0;
for (i=l;i<=k;i++) {
f=s*rv1[i];
rv1[i]=c*rv1[i];
if ((double)(fabs(f)+anorm) == anorm) break;
g=w[i];
h=PYTHAG(f,g);
w[i]=h;
h=1.0/h;
c=g*h;
s = -f*h;
for (j=0;j<m;j++) {
y=u[j][nm];
z=u[j][i];
u[j][nm]=y*c+z*s;
u[j][i]=z*c-y*s;
}
}
}
z=w[k];
if (l == k) {
if (z < 0.0) {
w[k] = -z;
for (j=0;j<n;j++) v[j][k] = -v[j][k];
}
break;
}
if (its == max_its - 1) {
return 1;
}
assert(k > 0);
x=w[l];
nm=k-1;
y=w[nm];
g=rv1[nm];
h=rv1[k];
f=((y-z)*(y+z)+(g-h)*(g+h))/(2.0*h*y);
g=PYTHAG(f,1.0);
f=((x-z)*(x+z)+h*((y/(f+SIGN(g,f)))-h))/x;
c=s=1.0;
for (j=l;j<=nm;j++) {
i=j+1;
g=rv1[i];
y=w[i];
h=s*g;
g=c*g;
z=PYTHAG(f,h);
rv1[j]=z;
c=f/z;
s=h/z;
f=x*c+g*s;
g = g*c-x*s;
h=y*s;
y *= c;
for (jj=0;jj<n;jj++) {
x=v[jj][j];
z=v[jj][i];
v[jj][j]=x*c+z*s;
v[jj][i]=z*c-x*s;
}
z=PYTHAG(f,h);
w[j]=z;
if (z) {
z=1.0/z;
c=f*z;
s=h*z;
}
f=c*g+s*y;
x=c*y-s*g;
for (jj=0;jj<m;jj++) {
y=u[jj][j];
z=u[jj][i];
u[jj][j]=y*c+z*s;
u[jj][i]=z*c-y*s;
}
}
rv1[l]=0.0;
rv1[k]=f;
w[k]=x;
}
}
free(rv1);
return 0;
}
static int SVD(double *U, double *W, double *V, double *matx, int M, int N) {
// Assumes allocation for U is MxN
double **nrU, **nrV;
int problem, i;
nrU = (double **)malloc((M)*sizeof(double*));
nrV = (double **)malloc((N)*sizeof(double*));
problem = !(nrU && nrV);
if (!problem) {
problem = 0;
for (i=0; i<M; i++) {
nrU[i] = &U[i * N];
}
for(i=0; i<N; i++) {
nrV[i] = &V[i * N];
}
}
if (problem) {
return 1;
}
/* copy from given matx into nrU */
for (i=0; i<M; i++) {
memcpy(&(nrU[i][0]), matx + N*i, N*sizeof(double));
}
/* HERE IT IS: do SVD */
if (svdcmp_(nrU, M, N, W, nrV)) {
return 1;
}
/* free Numerical Recipes arrays */
free(nrU);
free(nrV);
return 0;
}
int PseudoInverse(double *inv, double *matx, const int M, const int N) {
double *U, *W, *V, ans;
int i, j, k;
U = (double *)malloc(M*N*sizeof(double));
W = (double *)malloc(N*sizeof(double));
V = (double *)malloc(N*N*sizeof(double));
if (!(U && W && V)) {
return 1;
}
if (SVD(U, W, V, matx, M, N)) {
return 1;
}
for (i=0; i<N; i++) {
if (fabs(W[i]) < TINY_NEAR_ZERO) {
return 1;
}
}
for (i=0; i<N; i++) {
for (j=0; j<M; j++) {
ans = 0;
for (k=0; k<N; k++) {
ans += V[k + N*i]*U[k + N*j]/W[k];
}
inv[j + M*i] = ans;
}
}
free(U);
free(W);
free(V);
return 0;
}
static double compute_error(projectPointsType projectPoints,
double *points1, int stride1,
double *points2, int stride2,
int npoints, double *H, int *mask) {
int i, n = 0;
double pt[2];
double *mp1 = points1;
double *mp2 = points2;
double sqerr = 0.0;
if (projectPoints == NULL) return -1.0;
if (mask) {
for (i = 0; i < npoints; ++i, mp1+=stride1, mp2+=stride2) {
if (mask[i]) {
projectPoints(H, mp1, pt, 1, stride1, stride2);
sqerr += (pt[0] - mp2[0]) * (pt[0] - mp2[0]) +
(pt[1] - mp2[1]) * (pt[1] - mp2[1]);
n++;
}
}
} else {
for (i = 0; i < npoints; ++i, mp1+=stride1, mp2+=stride2) {
projectPoints(H, mp1, pt, 1, stride1, stride2);
sqerr += (pt[0] - mp2[0]) * (pt[0] - mp2[0]) +
(pt[1] - mp2[1]) * (pt[1] - mp2[1]);
n++;
}
}
return sqrt(sqerr / n);
}
////////////////////////////////////////////////////////////////////////////////
// ransac
typedef int (*isDegenerateType)(double *p);
typedef void (*normalizeType)(double *p, int np, double *T);
typedef void (*denormalizeType)(double *H, double *T1, double *T2);
typedef int (*findTransformationType)(int points,
double *points1,
double *points2,
double *H);
int ransac_(double *matched_points,
int npoints,
int *number_of_inliers,
int *best_inlier_mask,
double *bestH,
const int minpts,
const int paramdim,
isDegenerateType isDegenerate,
normalizeType normalize,
denormalizeType denormalize,
findTransformationType findTransformation,
projectPointsType projectPoints) {
static const double INLIER_THRESHOLD_NORMALIZED = 0.1;
static const double INLIER_THRESHOLD_UNNORMALIZED = 1.5;
static const double PROBABILITY_REQUIRED = 0.9;
static const double EPS = 1e-12;
static const int MIN_TRIALS = 20;
const double inlier_threshold = (normalize && denormalize ?
INLIER_THRESHOLD_NORMALIZED :
INLIER_THRESHOLD_UNNORMALIZED);
int N = 10000, trial_count = 0;
int i, j;
int max_inliers = 0;
double best_variance = 0.0;
double H[MAX_PARAMDIM];
double points1[2 * MAX_MINPTS];
double points2[2 * MAX_MINPTS];
double *best_inlier_set1;
double *best_inlier_set2;
double *inlier_set1;
double *inlier_set2;
double *corners1;
double *corners2;
double *image1_coord;
double *image2_coord;
int *inlier_mask;
double *cnp1, *cnp2;
double T1[9], T2[9];
srand((unsigned)time(NULL)) ;
// srand( 12345 ) ;
//
*number_of_inliers = 0;
if (npoints < minpts) {
printf("Cannot find motion with %d matches\n", npoints);
return 1;
}
best_inlier_set1 = (double *)malloc(sizeof(double) * npoints * 2);
best_inlier_set2 = (double *)malloc(sizeof(double) * npoints * 2);
inlier_set1 = (double *)malloc(sizeof(double) * npoints * 2);
inlier_set2 = (double *)malloc(sizeof(double) * npoints * 2);
corners1 = (double *)malloc(sizeof(double) * npoints * 2);
corners2 = (double *)malloc(sizeof(double) * npoints * 2);
image1_coord = (double *)malloc(sizeof(double) * npoints * 2);
image2_coord = (double *)malloc(sizeof(double) * npoints * 2);
inlier_mask = (int*)malloc(sizeof(int) * npoints);
for(cnp1 = corners1, cnp2 = corners2, i = 0; i < npoints; ++i) {
*(cnp1++) = *(matched_points++);
*(cnp1++) = *(matched_points++);
*(cnp2++) = *(matched_points++);
*(cnp2++) = *(matched_points++);
}
matched_points -= 4 * npoints;
if (normalize && denormalize) {
normalize(corners1, npoints, T1);
normalize(corners2, npoints, T2);
}
while (N > trial_count) {
int num_inliers = 0;
double sum_distance = 0.0;
double sum_distance_squared = 0.0;
int degenerate = 1;
while (degenerate) {
i = 0;
while (i < minpts) {
int index = rand() % npoints;
int duplicate = 0;
for (j = 0; j < i; ++j) {
if (points1[j*2] == corners1[index*2] &&
points1[j*2+1] == corners1[index*2+1]) {
duplicate = 1;
break;
}
}
if(duplicate) continue;
// add to list
points1[i*2] = corners1[index*2];
points1[i*2+1] = corners1[index*2+1];
points2[i*2] = corners2[index*2];
points2[i*2+1] = corners2[index*2+1];
i++;
}
degenerate = isDegenerate(points1);
}
if (findTransformation(minpts, points1, points2, H)) {
trial_count++;
continue;
}
projectPoints(H, corners1, image1_coord, npoints, 2, 2);
for( i = 0; i < npoints; ++i ) {
double dx = image1_coord[i*2] - corners2[i*2];
double dy = image1_coord[i*2 + 1] - corners2[i*2 + 1];
double distance = sqrt(dx*dx + dy*dy);
inlier_mask[i] = distance < inlier_threshold;
if (inlier_mask[i]) {
inlier_set1[num_inliers*2] = corners1[i*2];
inlier_set1[num_inliers*2 + 1] = corners1[i*2 + 1];
inlier_set2[num_inliers*2] = corners2[i*2];
inlier_set2[num_inliers*2 + 1] = corners2[i*2 + 1];
num_inliers++;
sum_distance += distance;
sum_distance_squared += distance*distance;
}
}
if (num_inliers >= max_inliers) {
double mean_distance = sum_distance / ((double)num_inliers);
double variance = sum_distance_squared / ((double)num_inliers - 1.0) -
mean_distance * mean_distance *
((double)num_inliers) / ((double)num_inliers - 1.0);
if ((num_inliers > max_inliers) ||
(num_inliers==max_inliers && variance < best_variance)) {
best_variance = variance;
max_inliers = num_inliers;
memcpy(bestH, H, paramdim * sizeof(double));
memcpy(best_inlier_set1, inlier_set1, num_inliers*2 * sizeof(double));
memcpy(best_inlier_set2, inlier_set2, num_inliers*2 * sizeof(double));
memcpy(best_inlier_mask, inlier_mask, npoints * sizeof(int));
if (num_inliers > 0) {
double fracinliers = (double)num_inliers/(double)npoints;
double pNoOutliers = 1 - pow(fracinliers, minpts);
int temp;
pNoOutliers = fmax(EPS, pNoOutliers);
pNoOutliers = fmin(1 - EPS, pNoOutliers);
temp = (int)(log(1.0 - PROBABILITY_REQUIRED)/log(pNoOutliers));
if (temp > 0 && temp < N) {
N = IMAX(temp, MIN_TRIALS);
}
}
}
}
trial_count++;
}
// printf("Number of trials = %d\n", trial_count);
findTransformation(max_inliers, best_inlier_set1, best_inlier_set2, bestH);
if (normalize && denormalize) {
denormalize(bestH, T1, T2);
}
*number_of_inliers = max_inliers;
/*
printf("Error score (all) = %g\n",
compute_error(projectPoints, matched_points, 4,
matched_points + 2, 4,
npoints, bestH, NULL));
printf("Error score (inliers) = %g\n",
compute_error(projectPoints, matched_points, 4,
matched_points + 2, 4,
npoints, bestH,
best_inlier_mask));
*/
free(best_inlier_set1);
free(best_inlier_set2);
free(inlier_set1);
free(inlier_set2);
free(corners1);
free(corners2);
free(image1_coord);
free(image2_coord);
free(inlier_mask);
return 0;
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////
static void normalizeHomography(double *pts, int n, double *T) {
// Assume the points are 2d coordinates with scale = 1
double *p = pts;
double mean[2] = {0, 0};
double msqe = 0;
double scale;
int i;
for (i = 0; i < n; ++i, p+=2) {
mean[0] += p[0];
mean[1] += p[1];
}
mean[0] /= n;
mean[1] /= n;
for (p = pts, i = 0; i < n; ++i, p+=2) {
p[0] -= mean[0];
p[1] -= mean[1];
msqe += sqrt(p[0] * p[0] + p[1] * p[1]);
}
msqe /= n;
scale = sqrt(2)/msqe;
T[0] = scale;
T[1] = 0;
T[2] = -scale * mean[0];
T[3] = 0;
T[4] = scale;
T[5] = -scale * mean[1];
T[6] = 0;
T[7] = 0;
T[8] = 1;
for (p = pts, i = 0; i < n; ++i, p+=2) {
p[0] *= scale;
p[1] *= scale;
}
}
static void invnormalize_mat(double *T, double *iT) {
double is = 1.0/T[0];
double m0 = -T[2]*is;
double m1 = -T[5]*is;
iT[0] = is;
iT[1] = 0;
iT[2] = m0;
iT[3] = 0;
iT[4] = is;
iT[5] = m1;
iT[6] = 0;
iT[7] = 0;
iT[8] = 1;
}
static void denormalizeHomography(double *H, double *T1, double *T2) {
double iT2[9];
double H2[9];
invnormalize_mat(T2, iT2);
MultiplyMat(H, T1, H2, 3, 3, 3);
MultiplyMat(iT2, H2, H, 3, 3, 3);
}
static void denormalizeAffine(double *H, double *T1, double *T2) {
double Ha[MAX_PARAMDIM];
Ha[0] = H[0];
Ha[1] = H[1];
Ha[2] = H[4];
Ha[3] = H[2];
Ha[4] = H[3];
Ha[5] = H[5];
Ha[6] = Ha[7] = 0;
Ha[8] = 1;
denormalizeHomography(Ha, T1, T2);
H[0] = Ha[0];
H[1] = Ha[1];
H[2] = Ha[3];
H[3] = Ha[4];
H[4] = Ha[2];
H[5] = Ha[5];
}
static void denormalizeRotZoom(double *H, double *T1, double *T2) {
double Ha[MAX_PARAMDIM];
memcpy(Ha, H, 6 * sizeof(*H));
Ha[0] = H[0];
Ha[1] = H[1];
Ha[2] = H[2];
Ha[3] = -H[1];
Ha[4] = H[0];
Ha[5] = H[3];
Ha[6] = Ha[7] = 0;
Ha[8] = 1;
denormalizeHomography(Ha, T1, T2);
H[0] = Ha[0];
H[1] = Ha[1];
H[2] = Ha[2];
H[3] = Ha[5];
}
static int is_collinear3(double *p1, double *p2, double *p3) {
static const double collinear_eps = 1e-3;
const double v = (p2[0] - p1[0]) * (p3[1] - p1[1]) - (p2[1] - p1[1]) * (p3[0] - p1[0]);
return fabs(v) < collinear_eps;
}
static int isDegenerateAffine(double *p) {
return is_collinear3(p, p + 2, p + 4);
}
static int isDegenerateHomography(double *p) {
return is_collinear3(p, p + 2, p + 4) ||
is_collinear3(p, p + 2, p + 6) ||
is_collinear3(p, p + 4, p + 6) ||
is_collinear3(p + 2, p + 4, p + 6);
}
void projectPointsRotZoom(double *mat, double *points, double *proj, const int n,
const int stride_points, const int stride_proj) {
int i;
for (i = 0; i < n; ++i) {
const double x = *(points++), y = *(points++);
*(proj++) = mat[0]*x + mat[1]*y + mat[2];
*(proj++) = -mat[1]*x + mat[0]*y + mat[3];
points += stride_points - 2;
proj += stride_proj - 2;
}
}
void projectPointsAffine(double *mat, double *points, double *proj, const int n,
const int stride_points, const int stride_proj) {
int i;
for (i = 0; i < n; ++i) {
const double x = *(points++), y = *(points++);
*(proj++) = mat[0]*x + mat[1]*y + mat[4];
*(proj++) = mat[2]*x + mat[3]*y + mat[5];
points += stride_points - 2;
proj += stride_proj - 2;
}
}
void projectPointsHomography(double *mat, double *points, double *proj, const int n,
const int stride_points, const int stride_proj) {
int i;
double x, y, Z;
for (i = 0; i < n; ++i) {
x = *(points++), y = *(points++);
Z = 1./(mat[6]*x + mat[7]*y + mat[8]);
*(proj++) = (mat[0]*x + mat[1]*y + mat[2])*Z;
*(proj++) = (mat[3]*x + mat[4]*y + mat[5])*Z;
points += stride_points - 2;
proj += stride_proj - 2;
}
}
int findRotZoom(const int np, double *pts1, double *pts2, double *mat) {
const int np2 = np * 2;
double *a = (double *)malloc(sizeof(double) * np2 * 9);
double *b = a + np2 * 4;
double *temp = b + np2;
int i;
double sx, sy, dx, dy;
double T1[9], T2[9];
normalizeHomography(pts1, np, T1);
normalizeHomography(pts2, np, T2);
for (i = 0; i < np; ++i) {
dx = *(pts2++);
dy = *(pts2++);
sx = *(pts1++);
sy = *(pts1++);
a[i * 2 * 4 + 0] = sx;
a[i * 2 * 4 + 1] = sy;
a[i * 2 * 4 + 2] = 1;
a[i * 2 * 4 + 3] = 0;
a[(i * 2 + 1) * 4 + 0] = sy;
a[(i * 2 + 1) * 4 + 1] = -sx;
a[(i * 2 + 1) * 4 + 2] = 0;
a[(i * 2 + 1) * 4 + 3] = 1;
b[2 * i] = dx;
b[2 * i + 1] = dy;
}
if (PseudoInverse(temp, a, np2, 4)){
free(a);
return 1;
}
MultiplyMat(temp, b, mat, 4, np2, 1);
denormalizeRotZoom(mat, T1, T2);
free(a);
return 0;
}
int findAffine(const int np, double *pts1, double *pts2, double *mat) {
const int np2 = np * 2;
double *a = (double *)malloc(sizeof(double) * np2 * 13);
double *b = a + np2 * 6;
double *temp = b + np2;
int i;
double sx, sy, dx, dy;
double T1[9], T2[9];
normalizeHomography(pts1, np, T1);
normalizeHomography(pts2, np, T2);
for (i = 0; i < np; ++i) {
dx = *(pts2++);
dy = *(pts2++);
sx = *(pts1++);
sy = *(pts1++);
a[i * 2 * 6 + 0] = sx;
a[i * 2 * 6 + 1] = sy;
a[i * 2 * 6 + 2] = 0;
a[i * 2 * 6 + 3] = 0;
a[i * 2 * 6 + 4] = 1;
a[i * 2 * 6 + 5] = 0;
a[(i * 2 + 1) * 6 + 0] = 0;
a[(i * 2 + 1) * 6 + 1] = 0;
a[(i * 2 + 1) * 6 + 2] = sx;
a[(i * 2 + 1) * 6 + 3] = sy;
a[(i * 2 + 1) * 6 + 4] = 0;
a[(i * 2 + 1) * 6 + 5] = 1;
b[2 * i] = dx;
b[2 * i + 1] = dy;
}
if (PseudoInverse(temp, a, np2, 6)){
free(a);
return 1;
}
MultiplyMat(temp, b, mat, 6, np2, 1);
denormalizeAffine(mat, T1, T2);
free(a);
return 0;
}
int findHomography(const int np, double *pts1, double *pts2, double *mat) {
// Implemented from Peter Kovesi's normalized implementation
const int np3 = np * 3;
double *a = (double *)malloc(sizeof(double) * np3 * 18);
double *U = a + np3 * 9;
double S[9], V[9 * 9];
int i, mini;
double sx, sy, dx, dy;
double T1[9], T2[9];
normalizeHomography(pts1, np, T1);
normalizeHomography(pts2, np, T2);
for (i = 0; i < np; ++i) {
dx = *(pts2++);
dy = *(pts2++);
sx = *(pts1++);
sy = *(pts1++);
a[i * 3 * 9 + 0] = a[i * 3 * 9 + 1] = a[i * 3 * 9 + 2] = 0;
a[i * 3 * 9 + 3] = -sx;
a[i * 3 * 9 + 4] = -sy;
a[i * 3 * 9 + 5] = -1;
a[i * 3 * 9 + 6] = dy * sx;
a[i * 3 * 9 + 7] = dy * sy;
a[i * 3 * 9 + 8] = dy;
a[(i * 3 + 1) * 9 + 0] = sx;
a[(i * 3 + 1) * 9 + 1] = sy;
a[(i * 3 + 1) * 9 + 2] = 1;
a[(i * 3 + 1) * 9 + 3] = a[(i * 3 + 1) * 9 + 4] =
a[(i * 3 + 1) * 9 + 5] = 0;
a[(i * 3 + 1) * 9 + 6] = -dx * sx;
a[(i * 3 + 1) * 9 + 7] = -dx * sy;
a[(i * 3 + 1) * 9 + 8] = -dx;
a[(i * 3 + 2) * 9 + 0] = -dy * sx;
a[(i * 3 + 2) * 9 + 1] = -dy * sy;
a[(i * 3 + 2) * 9 + 2] = -dy;
a[(i * 3 + 2) * 9 + 3] = dx * sx;
a[(i * 3 + 2) * 9 + 4] = dx * sy;
a[(i * 3 + 2) * 9 + 5] = dx;
a[(i * 3 + 2) * 9 + 6] = a[(i * 3 + 2) * 9 + 7] =
a[(i * 3 + 2) * 9 + 8] = 0;
}
if (SVD(U, S, V, a, np3, 9)) {
free(a);
return 1;
} else {
double minS = 1e12;
mini = -1;
for (i = 0; i < 9; ++i) {
if (S[i] < minS) {
minS = S[i];
mini = i;
}
}
}
for (i = 0; i < 9; i++) mat[i] = V[i * 9 + mini];
denormalizeHomography(mat, T1, T2);
// if (mat[8] == 0.0) return 1;
// for (i = 0; i < 9; i++) mat[i] /= mat[8];
free(a);
return 0;
}
int findHomographyScale1(const int np, double *pts1, double *pts2,
double *mat) {
// This implementation assumes h33 = 1, but does not seem to give good results
const int np2 = np * 2;
double *a = (double *)malloc(sizeof(double) * np2 * 17);
double *b = a + np2 * 8;
double *temp = b + np2;
int i, j;
double sx, sy, dx, dy;
double T1[9], T2[9];
normalizeHomography(pts1, np, T1);
normalizeHomography(pts2, np, T2);
for (i = 0, j = np; i < np; ++i, ++j) {
dx = *(pts2++);
dy = *(pts2++);
sx = *(pts1++);
sy = *(pts1++);
a[i * 8 + 0] = a[j * 8 + 3] = sx;
a[i * 8 + 1] = a[j * 8 + 4] = sy;
a[i * 8 + 2] = a[j * 8 + 5] = 1;
a[i * 8 + 3] = a[i * 8 + 4] = a[i * 8 + 5] =
a[j * 8 + 0] = a[j * 8 + 1] = a[j * 8 + 2] = 0;
a[i * 8 + 6] = -dx * sx;
a[i * 8 + 7] = -dx * sy;
a[j * 8 + 6] = -dy * sx;
a[j * 8 + 7] = -dy * sy;
b[i] = dx;
b[j] = dy;
}
if (PseudoInverse(temp, a, np2, 8)) {
free(a);
return 1;
}
MultiplyMat(temp, b, &*mat, 8, np2, 1);
mat[8] = 1;
denormalizeHomography(mat, T1, T2);
free(a);
return 0;
}
int ransacRotZoom(double *matched_points, int npoints,
int *number_of_inliers,
int *best_inlier_mask, double *bestH) {
return ransac_(matched_points,
npoints,
number_of_inliers,
best_inlier_mask,
bestH,
3,
4,
isDegenerateAffine,
NULL, // normalizeHomography,
NULL, // denormalizeRotZoom,
findRotZoom,
projectPointsRotZoom);
}
int ransacAffine(double *matched_points, int npoints,
int *number_of_inliers,
int *best_inlier_mask, double *bestH) {
return ransac_(matched_points,
npoints,
number_of_inliers,
best_inlier_mask,
bestH,
3,
6,
isDegenerateAffine,
NULL, // normalizeHomography,
NULL, // denormalizeAffine,
findAffine,
projectPointsAffine);
}
int ransacHomography(double *matched_points, int npoints,
int *number_of_inliers,
int *best_inlier_mask, double *bestH) {
return ransac_(matched_points,
npoints,
number_of_inliers,
best_inlier_mask,
bestH,
4,
8,
isDegenerateHomography,
NULL, // normalizeHomography,
NULL, // denormalizeHomography,
findHomography,
projectPointsHomography);
}

43
vp9/encoder/vp9_ransac.h Normal file
View File

@ -0,0 +1,43 @@
/*
* Copyright (c) 2015 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be
* found in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef VP9_ENCODER_VP9_RANSAC_H
#define VP9_ENCODER_VP9_RANSAC_H
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <memory.h>
typedef int (*ransacType)(double *matched_points, int npoints,
int *number_of_inliers, int *best_inlier_mask,
double *bestH);
typedef void (*projectPointsType)(double *mat, double *points, double *proj,
const int n, const int stride_points,
const int stride_proj);
int ransacHomography(double *matched_points, int npoints,
int *number_of_inliers, int *best_inlier_indices,
double *bestH);
int ransacAffine(double *matched_points, int npoints,
int *number_of_inliers, int *best_inlier_indices,
double *bestH);
int ransacRotZoom(double *matched_points, int npoints,
int *number_of_inliers, int *best_inlier_indices,
double *bestH);
void projectPointsHomography(double *mat, double *points, double *proj,
const int n, const int stride_points, const int stride_proj);
void projectPointsAffine(double *mat, double *points, double *proj,
const int n, const int stride_points, const int stride_proj);
void projectPointsRotZoom(double *mat, double *points, double *proj,
const int n, const int stride_points, const int stride_proj);
#endif // VP9_ENCODER_VP9_RANSAC_H

View File

@ -74,6 +74,14 @@ VP9_CX_SRCS-yes += encoder/vp9_subexp.h
VP9_CX_SRCS-yes += encoder/vp9_svc_layercontext.c
VP9_CX_SRCS-yes += encoder/vp9_resize.c
VP9_CX_SRCS-yes += encoder/vp9_resize.h
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_ransac.c
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_ransac.h
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_corner_detect.c
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_corner_detect.h
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_corner_match.c
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_corner_match.h
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_global_motion.c
VP9_CX_SRCS-$(CONFIG_EXPERIMENTAL) += encoder/vp9_global_motion.h
VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_ssim.c
VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_ssim.h
VP9_CX_SRCS-yes += encoder/vp9_tokenize.c