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 <stdio.h>
13 #include <stdlib.h>
14 #include <memory.h>
15 #include <math.h>
16 #include <assert.h>
17
18 #include "av1/encoder/global_motion.h"
19
20 #include "av1/common/warped_motion.h"
21
22 #include "av1/encoder/segmentation.h"
23 #include "av1/encoder/corner_detect.h"
24 #include "av1/encoder/corner_match.h"
25 #include "av1/encoder/ransac.h"
26
27 #define MAX_CORNERS 4096
28 #define MIN_INLIER_PROB 0.1
29
30 #define MIN_TRANS_THRESH (1 * GM_TRANS_DECODE_FACTOR)
31
32 // Border over which to compute the global motion
33 #define ERRORADV_BORDER 0
34
35 #define ERRORADV_MAX_THRESH 0.995
36 #define ERRORADV_COST_PRODUCT_THRESH 26000
37
is_enough_erroradvantage(double best_erroradvantage,int params_cost)38 int is_enough_erroradvantage(double best_erroradvantage, int params_cost) {
39 return best_erroradvantage < ERRORADV_MAX_THRESH &&
40 best_erroradvantage * params_cost < ERRORADV_COST_PRODUCT_THRESH;
41 }
42
convert_to_params(const double * params,int32_t * model)43 static void convert_to_params(const double *params, int32_t *model) {
44 int i;
45 int alpha_present = 0;
46 model[0] = (int32_t)floor(params[0] * (1 << GM_TRANS_PREC_BITS) + 0.5);
47 model[1] = (int32_t)floor(params[1] * (1 << GM_TRANS_PREC_BITS) + 0.5);
48 model[0] = (int32_t)clamp(model[0], GM_TRANS_MIN, GM_TRANS_MAX) *
49 GM_TRANS_DECODE_FACTOR;
50 model[1] = (int32_t)clamp(model[1], GM_TRANS_MIN, GM_TRANS_MAX) *
51 GM_TRANS_DECODE_FACTOR;
52
53 for (i = 2; i < 6; ++i) {
54 const int diag_value = ((i == 2 || i == 5) ? (1 << GM_ALPHA_PREC_BITS) : 0);
55 model[i] = (int32_t)floor(params[i] * (1 << GM_ALPHA_PREC_BITS) + 0.5);
56 model[i] =
57 (int32_t)clamp(model[i] - diag_value, GM_ALPHA_MIN, GM_ALPHA_MAX);
58 alpha_present |= (model[i] != 0);
59 model[i] = (model[i] + diag_value) * GM_ALPHA_DECODE_FACTOR;
60 }
61 for (; i < 8; ++i) {
62 model[i] = (int32_t)floor(params[i] * (1 << GM_ROW3HOMO_PREC_BITS) + 0.5);
63 model[i] = (int32_t)clamp(model[i], GM_ROW3HOMO_MIN, GM_ROW3HOMO_MAX) *
64 GM_ROW3HOMO_DECODE_FACTOR;
65 alpha_present |= (model[i] != 0);
66 }
67
68 if (!alpha_present) {
69 if (abs(model[0]) < MIN_TRANS_THRESH && abs(model[1]) < MIN_TRANS_THRESH) {
70 model[0] = 0;
71 model[1] = 0;
72 }
73 }
74 }
75
convert_model_to_params(const double * params,WarpedMotionParams * model)76 void convert_model_to_params(const double *params, WarpedMotionParams *model) {
77 convert_to_params(params, model->wmmat);
78 model->wmtype = get_gmtype(model);
79 }
80
81 // Adds some offset to a global motion parameter and handles
82 // all of the necessary precision shifts, clamping, and
83 // zero-centering.
add_param_offset(int param_index,int32_t param_value,int32_t offset)84 static int32_t add_param_offset(int param_index, int32_t param_value,
85 int32_t offset) {
86 const int scale_vals[3] = { GM_TRANS_PREC_DIFF, GM_ALPHA_PREC_DIFF,
87 GM_ROW3HOMO_PREC_DIFF };
88 const int clamp_vals[3] = { GM_TRANS_MAX, GM_ALPHA_MAX, GM_ROW3HOMO_MAX };
89 // type of param: 0 - translation, 1 - affine, 2 - homography
90 const int param_type = (param_index < 2 ? 0 : (param_index < 6 ? 1 : 2));
91 const int is_one_centered = (param_index == 2 || param_index == 5);
92
93 // Make parameter zero-centered and offset the shift that was done to make
94 // it compatible with the warped model
95 param_value = (param_value - (is_one_centered << WARPEDMODEL_PREC_BITS)) >>
96 scale_vals[param_type];
97 // Add desired offset to the rescaled/zero-centered parameter
98 param_value += offset;
99 // Clamp the parameter so it does not overflow the number of bits allotted
100 // to it in the bitstream
101 param_value = (int32_t)clamp(param_value, -clamp_vals[param_type],
102 clamp_vals[param_type]);
103 // Rescale the parameter to WARPEDMODEL_PRECISION_BITS so it is compatible
104 // with the warped motion library
105 param_value *= (1 << scale_vals[param_type]);
106
107 // Undo the zero-centering step if necessary
108 return param_value + (is_one_centered << WARPEDMODEL_PREC_BITS);
109 }
110
force_wmtype(WarpedMotionParams * wm,TransformationType wmtype)111 static void force_wmtype(WarpedMotionParams *wm, TransformationType wmtype) {
112 switch (wmtype) {
113 case IDENTITY: wm->wmmat[0] = 0; wm->wmmat[1] = 0;
114 case TRANSLATION:
115 wm->wmmat[2] = 1 << WARPEDMODEL_PREC_BITS;
116 wm->wmmat[3] = 0;
117 case ROTZOOM: wm->wmmat[4] = -wm->wmmat[3]; wm->wmmat[5] = wm->wmmat[2];
118 case AFFINE: wm->wmmat[6] = wm->wmmat[7] = 0; break;
119 case HORTRAPEZOID: wm->wmmat[6] = wm->wmmat[4] = 0; break;
120 case VERTRAPEZOID: wm->wmmat[7] = wm->wmmat[3] = 0; break;
121 case HOMOGRAPHY: break;
122 default: assert(0);
123 }
124 wm->wmtype = wmtype;
125 }
126
refine_integerized_param(WarpedMotionParams * wm,TransformationType wmtype,int use_hbd,int bd,uint8_t * ref,int r_width,int r_height,int r_stride,uint8_t * dst,int d_width,int d_height,int d_stride,int n_refinements,int64_t best_frame_error)127 int64_t refine_integerized_param(WarpedMotionParams *wm,
128 TransformationType wmtype,
129 #if CONFIG_HIGHBITDEPTH
130 int use_hbd, int bd,
131 #endif // CONFIG_HIGHBITDEPTH
132 uint8_t *ref, int r_width, int r_height,
133 int r_stride, uint8_t *dst, int d_width,
134 int d_height, int d_stride, int n_refinements,
135 int64_t best_frame_error) {
136 static const int max_trans_model_params[TRANS_TYPES] = {
137 0, 2, 4, 6, 8, 8, 8
138 };
139 const int border = ERRORADV_BORDER;
140 int i = 0, p;
141 int n_params = max_trans_model_params[wmtype];
142 int32_t *param_mat = wm->wmmat;
143 int64_t step_error, best_error;
144 int32_t step;
145 int32_t *param;
146 int32_t curr_param;
147 int32_t best_param;
148
149 force_wmtype(wm, wmtype);
150 best_error = av1_warp_error(
151 wm,
152 #if CONFIG_HIGHBITDEPTH
153 use_hbd, bd,
154 #endif // CONFIG_HIGHBITDEPTH
155 ref, r_width, r_height, r_stride, dst + border * d_stride + border,
156 border, border, d_width - 2 * border, d_height - 2 * border, d_stride, 0,
157 0, SCALE_SUBPEL_SHIFTS, SCALE_SUBPEL_SHIFTS, best_frame_error);
158 best_error = AOMMIN(best_error, best_frame_error);
159 step = 1 << (n_refinements - 1);
160 for (i = 0; i < n_refinements; i++, step >>= 1) {
161 for (p = 0; p < n_params; ++p) {
162 int step_dir = 0;
163 // Skip searches for parameters that are forced to be 0
164 if (wmtype == HORTRAPEZOID && (p == 4 || p == 6)) continue;
165 if (wmtype == VERTRAPEZOID && (p == 3 || p == 7)) continue;
166 param = param_mat + p;
167 curr_param = *param;
168 best_param = curr_param;
169 // look to the left
170 *param = add_param_offset(p, curr_param, -step);
171 step_error = av1_warp_error(
172 wm,
173 #if CONFIG_HIGHBITDEPTH
174 use_hbd, bd,
175 #endif // CONFIG_HIGHBITDEPTH
176 ref, r_width, r_height, r_stride, dst + border * d_stride + border,
177 border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
178 0, 0, SCALE_SUBPEL_SHIFTS, SCALE_SUBPEL_SHIFTS, best_error);
179 if (step_error < best_error) {
180 best_error = step_error;
181 best_param = *param;
182 step_dir = -1;
183 }
184
185 // look to the right
186 *param = add_param_offset(p, curr_param, step);
187 step_error = av1_warp_error(
188 wm,
189 #if CONFIG_HIGHBITDEPTH
190 use_hbd, bd,
191 #endif // CONFIG_HIGHBITDEPTH
192 ref, r_width, r_height, r_stride, dst + border * d_stride + border,
193 border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
194 0, 0, SCALE_SUBPEL_SHIFTS, SCALE_SUBPEL_SHIFTS, best_error);
195 if (step_error < best_error) {
196 best_error = step_error;
197 best_param = *param;
198 step_dir = 1;
199 }
200 *param = best_param;
201
202 // look to the direction chosen above repeatedly until error increases
203 // for the biggest step size
204 while (step_dir) {
205 *param = add_param_offset(p, best_param, step * step_dir);
206 step_error = av1_warp_error(
207 wm,
208 #if CONFIG_HIGHBITDEPTH
209 use_hbd, bd,
210 #endif // CONFIG_HIGHBITDEPTH
211 ref, r_width, r_height, r_stride, dst + border * d_stride + border,
212 border, border, d_width - 2 * border, d_height - 2 * border,
213 d_stride, 0, 0, SCALE_SUBPEL_SHIFTS, SCALE_SUBPEL_SHIFTS,
214 best_error);
215 if (step_error < best_error) {
216 best_error = step_error;
217 best_param = *param;
218 } else {
219 *param = best_param;
220 step_dir = 0;
221 }
222 }
223 }
224 }
225 force_wmtype(wm, wmtype);
226 wm->wmtype = get_gmtype(wm);
227 return best_error;
228 }
229
get_ransac_type(TransformationType type)230 static INLINE RansacFunc get_ransac_type(TransformationType type) {
231 switch (type) {
232 case HOMOGRAPHY: return ransac_homography;
233 case HORTRAPEZOID: return ransac_hortrapezoid;
234 case VERTRAPEZOID: return ransac_vertrapezoid;
235 case AFFINE: return ransac_affine;
236 case ROTZOOM: return ransac_rotzoom;
237 case TRANSLATION: return ransac_translation;
238 default: assert(0); return NULL;
239 }
240 }
241
242 #if CONFIG_HIGHBITDEPTH
downconvert_frame(YV12_BUFFER_CONFIG * frm,int bit_depth)243 static unsigned char *downconvert_frame(YV12_BUFFER_CONFIG *frm,
244 int bit_depth) {
245 int i, j;
246 uint16_t *orig_buf = CONVERT_TO_SHORTPTR(frm->y_buffer);
247 uint8_t *buf_8bit = frm->y_buffer_8bit;
248 assert(buf_8bit);
249 if (!frm->buf_8bit_valid) {
250 for (i = 0; i < frm->y_height; ++i) {
251 for (j = 0; j < frm->y_width; ++j) {
252 buf_8bit[i * frm->y_stride + j] =
253 orig_buf[i * frm->y_stride + j] >> (bit_depth - 8);
254 }
255 }
256 frm->buf_8bit_valid = 1;
257 }
258 return buf_8bit;
259 }
260 #endif
261
compute_global_motion_feature_based(TransformationType type,YV12_BUFFER_CONFIG * frm,YV12_BUFFER_CONFIG * ref,int bit_depth,int * num_inliers_by_motion,double * params_by_motion,int num_motions)262 int compute_global_motion_feature_based(
263 TransformationType type, YV12_BUFFER_CONFIG *frm, YV12_BUFFER_CONFIG *ref,
264 #if CONFIG_HIGHBITDEPTH
265 int bit_depth,
266 #endif
267 int *num_inliers_by_motion, double *params_by_motion, int num_motions) {
268 int i;
269 int num_frm_corners, num_ref_corners;
270 int num_correspondences;
271 int *correspondences;
272 int frm_corners[2 * MAX_CORNERS], ref_corners[2 * MAX_CORNERS];
273 unsigned char *frm_buffer = frm->y_buffer;
274 unsigned char *ref_buffer = ref->y_buffer;
275 RansacFunc ransac = get_ransac_type(type);
276
277 #if CONFIG_HIGHBITDEPTH
278 if (frm->flags & YV12_FLAG_HIGHBITDEPTH) {
279 // The frame buffer is 16-bit, so we need to convert to 8 bits for the
280 // following code. We cache the result until the frame is released.
281 frm_buffer = downconvert_frame(frm, bit_depth);
282 }
283 if (ref->flags & YV12_FLAG_HIGHBITDEPTH) {
284 ref_buffer = downconvert_frame(ref, bit_depth);
285 }
286 #endif
287
288 // compute interest points in images using FAST features
289 num_frm_corners = fast_corner_detect(frm_buffer, frm->y_width, frm->y_height,
290 frm->y_stride, frm_corners, MAX_CORNERS);
291 num_ref_corners = fast_corner_detect(ref_buffer, ref->y_width, ref->y_height,
292 ref->y_stride, ref_corners, MAX_CORNERS);
293
294 // find correspondences between the two images
295 correspondences =
296 (int *)malloc(num_frm_corners * 4 * sizeof(*correspondences));
297 num_correspondences = determine_correspondence(
298 frm_buffer, (int *)frm_corners, num_frm_corners, ref_buffer,
299 (int *)ref_corners, num_ref_corners, frm->y_width, frm->y_height,
300 frm->y_stride, ref->y_stride, correspondences);
301
302 ransac(correspondences, num_correspondences, num_inliers_by_motion,
303 params_by_motion, num_motions);
304
305 free(correspondences);
306
307 // Set num_inliers = 0 for motions with too few inliers so they are ignored.
308 for (i = 0; i < num_motions; ++i) {
309 if (num_inliers_by_motion[i] < MIN_INLIER_PROB * num_correspondences) {
310 num_inliers_by_motion[i] = 0;
311 }
312 }
313
314 // Return true if any one of the motions has inliers.
315 for (i = 0; i < num_motions; ++i) {
316 if (num_inliers_by_motion[i] > 0) return 1;
317 }
318 return 0;
319 }
320