1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <stdlib.h>
13 #include <memory.h>
14 #include <math.h>
15
16 #include "config/av1_rtcd.h"
17
18 #include "aom_ports/system_state.h"
19 #include "av1/encoder/corner_match.h"
20
21 #define SEARCH_SZ 9
22 #define SEARCH_SZ_BY2 ((SEARCH_SZ - 1) / 2)
23
24 #define THRESHOLD_NCC 0.75
25
26 /* Compute var(im) * MATCH_SZ_SQ over a MATCH_SZ by MATCH_SZ window of im,
27 centered at (x, y).
28 */
compute_variance(unsigned char * im,int stride,int x,int y)29 static double compute_variance(unsigned char *im, int stride, int x, int y) {
30 int sum = 0;
31 int sumsq = 0;
32 int var;
33 int i, j;
34 for (i = 0; i < MATCH_SZ; ++i)
35 for (j = 0; j < MATCH_SZ; ++j) {
36 sum += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
37 sumsq += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)] *
38 im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
39 }
40 var = sumsq * MATCH_SZ_SQ - sum * sum;
41 return (double)var;
42 }
43
44 /* Compute corr(im1, im2) * MATCH_SZ * stddev(im1), where the
45 correlation/standard deviation are taken over MATCH_SZ by MATCH_SZ windows
46 of each image, centered at (x1, y1) and (x2, y2) respectively.
47 */
av1_compute_cross_correlation_c(unsigned char * im1,int stride1,int x1,int y1,unsigned char * im2,int stride2,int x2,int y2)48 double av1_compute_cross_correlation_c(unsigned char *im1, int stride1, int x1,
49 int y1, unsigned char *im2, int stride2,
50 int x2, int y2) {
51 int v1, v2;
52 int sum1 = 0;
53 int sum2 = 0;
54 int sumsq2 = 0;
55 int cross = 0;
56 int var2, cov;
57 int i, j;
58 for (i = 0; i < MATCH_SZ; ++i)
59 for (j = 0; j < MATCH_SZ; ++j) {
60 v1 = im1[(i + y1 - MATCH_SZ_BY2) * stride1 + (j + x1 - MATCH_SZ_BY2)];
61 v2 = im2[(i + y2 - MATCH_SZ_BY2) * stride2 + (j + x2 - MATCH_SZ_BY2)];
62 sum1 += v1;
63 sum2 += v2;
64 sumsq2 += v2 * v2;
65 cross += v1 * v2;
66 }
67 var2 = sumsq2 * MATCH_SZ_SQ - sum2 * sum2;
68 cov = cross * MATCH_SZ_SQ - sum1 * sum2;
69 aom_clear_system_state();
70 return cov / sqrt((double)var2);
71 }
72
is_eligible_point(int pointx,int pointy,int width,int height)73 static int is_eligible_point(int pointx, int pointy, int width, int height) {
74 return (pointx >= MATCH_SZ_BY2 && pointy >= MATCH_SZ_BY2 &&
75 pointx + MATCH_SZ_BY2 < width && pointy + MATCH_SZ_BY2 < height);
76 }
77
is_eligible_distance(int point1x,int point1y,int point2x,int point2y,int width,int height)78 static int is_eligible_distance(int point1x, int point1y, int point2x,
79 int point2y, int width, int height) {
80 const int thresh = (width < height ? height : width) >> 4;
81 return ((point1x - point2x) * (point1x - point2x) +
82 (point1y - point2y) * (point1y - point2y)) <= thresh * thresh;
83 }
84
improve_correspondence(unsigned char * frm,unsigned char * ref,int width,int height,int frm_stride,int ref_stride,Correspondence * correspondences,int num_correspondences)85 static void improve_correspondence(unsigned char *frm, unsigned char *ref,
86 int width, int height, int frm_stride,
87 int ref_stride,
88 Correspondence *correspondences,
89 int num_correspondences) {
90 int i;
91 for (i = 0; i < num_correspondences; ++i) {
92 int x, y, best_x = 0, best_y = 0;
93 double best_match_ncc = 0.0;
94 for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y) {
95 for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
96 double match_ncc;
97 if (!is_eligible_point(correspondences[i].rx + x,
98 correspondences[i].ry + y, width, height))
99 continue;
100 if (!is_eligible_distance(correspondences[i].x, correspondences[i].y,
101 correspondences[i].rx + x,
102 correspondences[i].ry + y, width, height))
103 continue;
104 match_ncc = av1_compute_cross_correlation(
105 frm, frm_stride, correspondences[i].x, correspondences[i].y, ref,
106 ref_stride, correspondences[i].rx + x, correspondences[i].ry + y);
107 if (match_ncc > best_match_ncc) {
108 best_match_ncc = match_ncc;
109 best_y = y;
110 best_x = x;
111 }
112 }
113 }
114 correspondences[i].rx += best_x;
115 correspondences[i].ry += best_y;
116 }
117 for (i = 0; i < num_correspondences; ++i) {
118 int x, y, best_x = 0, best_y = 0;
119 double best_match_ncc = 0.0;
120 for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y)
121 for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
122 double match_ncc;
123 if (!is_eligible_point(correspondences[i].x + x,
124 correspondences[i].y + y, width, height))
125 continue;
126 if (!is_eligible_distance(
127 correspondences[i].x + x, correspondences[i].y + y,
128 correspondences[i].rx, correspondences[i].ry, width, height))
129 continue;
130 match_ncc = av1_compute_cross_correlation(
131 ref, ref_stride, correspondences[i].rx, correspondences[i].ry, frm,
132 frm_stride, correspondences[i].x + x, correspondences[i].y + y);
133 if (match_ncc > best_match_ncc) {
134 best_match_ncc = match_ncc;
135 best_y = y;
136 best_x = x;
137 }
138 }
139 correspondences[i].x += best_x;
140 correspondences[i].y += best_y;
141 }
142 }
143
av1_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)144 int av1_determine_correspondence(unsigned char *frm, int *frm_corners,
145 int num_frm_corners, unsigned char *ref,
146 int *ref_corners, int num_ref_corners,
147 int width, int height, int frm_stride,
148 int ref_stride, int *correspondence_pts) {
149 // TODO(sarahparker) Improve this to include 2-way match
150 int i, j;
151 Correspondence *correspondences = (Correspondence *)correspondence_pts;
152 int num_correspondences = 0;
153 for (i = 0; i < num_frm_corners; ++i) {
154 double best_match_ncc = 0.0;
155 double template_norm;
156 int best_match_j = -1;
157 if (!is_eligible_point(frm_corners[2 * i], frm_corners[2 * i + 1], width,
158 height))
159 continue;
160 for (j = 0; j < num_ref_corners; ++j) {
161 double match_ncc;
162 if (!is_eligible_point(ref_corners[2 * j], ref_corners[2 * j + 1], width,
163 height))
164 continue;
165 if (!is_eligible_distance(frm_corners[2 * i], frm_corners[2 * i + 1],
166 ref_corners[2 * j], ref_corners[2 * j + 1],
167 width, height))
168 continue;
169 match_ncc = av1_compute_cross_correlation(
170 frm, frm_stride, frm_corners[2 * i], frm_corners[2 * i + 1], ref,
171 ref_stride, ref_corners[2 * j], ref_corners[2 * j + 1]);
172 if (match_ncc > best_match_ncc) {
173 best_match_ncc = match_ncc;
174 best_match_j = j;
175 }
176 }
177 // Note: We want to test if the best correlation is >= THRESHOLD_NCC,
178 // but need to account for the normalization in
179 // av1_compute_cross_correlation.
180 template_norm = compute_variance(frm, frm_stride, frm_corners[2 * i],
181 frm_corners[2 * i + 1]);
182 if (best_match_ncc > THRESHOLD_NCC * sqrt(template_norm)) {
183 correspondences[num_correspondences].x = frm_corners[2 * i];
184 correspondences[num_correspondences].y = frm_corners[2 * i + 1];
185 correspondences[num_correspondences].rx = ref_corners[2 * best_match_j];
186 correspondences[num_correspondences].ry =
187 ref_corners[2 * best_match_j + 1];
188 num_correspondences++;
189 }
190 }
191 improve_correspondence(frm, ref, width, height, frm_stride, ref_stride,
192 correspondences, num_correspondences);
193 return num_correspondences;
194 }
195