1 /**
2 * Copyright (c) 2016 Davinder Singh (DSM_) <ds.mudhar<@gmail.com>
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 #include "motion_estimation.h"
22 #include "libavcodec/mathops.h"
23 #include "libavutil/avassert.h"
24 #include "libavutil/common.h"
25 #include "libavutil/imgutils.h"
26 #include "libavutil/opt.h"
27 #include "libavutil/pixdesc.h"
28 #include "libavutil/motion_vector.h"
29 #include "avfilter.h"
30 #include "formats.h"
31 #include "internal.h"
32 #include "video.h"
33
34 typedef struct MEContext {
35 const AVClass *class;
36 AVMotionEstContext me_ctx;
37 int method; ///< motion estimation method
38
39 int mb_size; ///< macroblock size
40 int search_param; ///< search parameter
41 int b_width, b_height, b_count;
42 int log2_mb_size;
43
44 AVFrame *prev, *cur, *next;
45
46 int (*mv_table[3])[2][2]; ///< motion vectors of current & prev 2 frames
47 } MEContext;
48
49 #define OFFSET(x) offsetof(MEContext, x)
50 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
51 #define CONST(name, help, val, unit) { name, help, 0, AV_OPT_TYPE_CONST, {.i64=val}, 0, 0, FLAGS, unit }
52
53 static const AVOption mestimate_options[] = {
54 { "method", "motion estimation method", OFFSET(method), AV_OPT_TYPE_INT, {.i64 = AV_ME_METHOD_ESA}, AV_ME_METHOD_ESA, AV_ME_METHOD_UMH, FLAGS, "method" },
55 CONST("esa", "exhaustive search", AV_ME_METHOD_ESA, "method"),
56 CONST("tss", "three step search", AV_ME_METHOD_TSS, "method"),
57 CONST("tdls", "two dimensional logarithmic search", AV_ME_METHOD_TDLS, "method"),
58 CONST("ntss", "new three step search", AV_ME_METHOD_NTSS, "method"),
59 CONST("fss", "four step search", AV_ME_METHOD_FSS, "method"),
60 CONST("ds", "diamond search", AV_ME_METHOD_DS, "method"),
61 CONST("hexbs", "hexagon-based search", AV_ME_METHOD_HEXBS, "method"),
62 CONST("epzs", "enhanced predictive zonal search", AV_ME_METHOD_EPZS, "method"),
63 CONST("umh", "uneven multi-hexagon search", AV_ME_METHOD_UMH, "method"),
64 { "mb_size", "macroblock size", OFFSET(mb_size), AV_OPT_TYPE_INT, {.i64 = 16}, 8, INT_MAX, FLAGS },
65 { "search_param", "search parameter", OFFSET(search_param), AV_OPT_TYPE_INT, {.i64 = 7}, 4, INT_MAX, FLAGS },
66 { NULL }
67 };
68
69 AVFILTER_DEFINE_CLASS(mestimate);
70
query_formats(AVFilterContext * ctx)71 static int query_formats(AVFilterContext *ctx)
72 {
73 static const enum AVPixelFormat pix_fmts[] = {
74 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
75 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
76 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
77 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
78 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
79 AV_PIX_FMT_YUVJ411P,
80 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
81 AV_PIX_FMT_GRAY8,
82 AV_PIX_FMT_NONE
83 };
84
85 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
86 if (!fmts_list)
87 return AVERROR(ENOMEM);
88 return ff_set_common_formats(ctx, fmts_list);
89 }
90
config_input(AVFilterLink * inlink)91 static int config_input(AVFilterLink *inlink)
92 {
93 MEContext *s = inlink->dst->priv;
94 int i;
95
96 s->log2_mb_size = av_ceil_log2_c(s->mb_size);
97 s->mb_size = 1 << s->log2_mb_size;
98
99 s->b_width = inlink->w >> s->log2_mb_size;
100 s->b_height = inlink->h >> s->log2_mb_size;
101 s->b_count = s->b_width * s->b_height;
102
103 if (s->b_count == 0)
104 return AVERROR(EINVAL);
105
106 for (i = 0; i < 3; i++) {
107 s->mv_table[i] = av_mallocz_array(s->b_count, sizeof(*s->mv_table[0]));
108 if (!s->mv_table[i])
109 return AVERROR(ENOMEM);
110 }
111
112 ff_me_init_context(&s->me_ctx, s->mb_size, s->search_param, inlink->w, inlink->h, 0, (s->b_width - 1) << s->log2_mb_size, 0, (s->b_height - 1) << s->log2_mb_size);
113
114 return 0;
115 }
116
add_mv_data(AVMotionVector * mv,int mb_size,int x,int y,int x_mv,int y_mv,int dir)117 static void add_mv_data(AVMotionVector *mv, int mb_size,
118 int x, int y, int x_mv, int y_mv, int dir)
119 {
120 mv->w = mb_size;
121 mv->h = mb_size;
122 mv->dst_x = x + (mb_size >> 1);
123 mv->dst_y = y + (mb_size >> 1);
124 mv->src_x = x_mv + (mb_size >> 1);
125 mv->src_y = y_mv + (mb_size >> 1);
126 mv->source = dir ? 1 : -1;
127 mv->flags = 0;
128 }
129
130 #define SEARCH_MV(method)\
131 do {\
132 for (mb_y = 0; mb_y < s->b_height; mb_y++)\
133 for (mb_x = 0; mb_x < s->b_width; mb_x++) {\
134 const int x_mb = mb_x << s->log2_mb_size;\
135 const int y_mb = mb_y << s->log2_mb_size;\
136 int mv[2] = {x_mb, y_mb};\
137 ff_me_search_##method(me_ctx, x_mb, y_mb, mv);\
138 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, me_ctx->mb_size, x_mb, y_mb, mv[0], mv[1], dir);\
139 }\
140 } while (0)
141
142 #define ADD_PRED(preds, px, py)\
143 do {\
144 preds.mvs[preds.nb][0] = px;\
145 preds.mvs[preds.nb][1] = py;\
146 preds.nb++;\
147 } while(0)
148
filter_frame(AVFilterLink * inlink,AVFrame * frame)149 static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
150 {
151 AVFilterContext *ctx = inlink->dst;
152 MEContext *s = ctx->priv;
153 AVMotionEstContext *me_ctx = &s->me_ctx;
154 AVFrameSideData *sd;
155 AVFrame *out;
156 int mb_x, mb_y, dir;
157 int32_t mv_count = 0;
158 int ret;
159
160 if (frame->pts == AV_NOPTS_VALUE) {
161 ret = ff_filter_frame(ctx->outputs[0], frame);
162 return ret;
163 }
164
165 av_frame_free(&s->prev);
166 s->prev = s->cur;
167 s->cur = s->next;
168 s->next = frame;
169
170 s->mv_table[2] = memcpy(s->mv_table[2], s->mv_table[1], sizeof(*s->mv_table[1]) * s->b_count);
171 s->mv_table[1] = memcpy(s->mv_table[1], s->mv_table[0], sizeof(*s->mv_table[0]) * s->b_count);
172
173 if (!s->cur) {
174 s->cur = av_frame_clone(frame);
175 if (!s->cur)
176 return AVERROR(ENOMEM);
177 }
178
179 if (!s->prev)
180 return 0;
181
182 out = av_frame_clone(s->cur);
183 if (!out)
184 return AVERROR(ENOMEM);
185
186 sd = av_frame_new_side_data(out, AV_FRAME_DATA_MOTION_VECTORS, 2 * s->b_count * sizeof(AVMotionVector));
187 if (!sd) {
188 av_frame_free(&out);
189 return AVERROR(ENOMEM);
190 }
191
192 me_ctx->data_cur = s->cur->data[0];
193 me_ctx->linesize = s->cur->linesize[0];
194
195 for (dir = 0; dir < 2; dir++) {
196 me_ctx->data_ref = (dir ? s->next : s->prev)->data[0];
197
198 if (s->method == AV_ME_METHOD_DS)
199 SEARCH_MV(ds);
200 else if (s->method == AV_ME_METHOD_ESA)
201 SEARCH_MV(esa);
202 else if (s->method == AV_ME_METHOD_FSS)
203 SEARCH_MV(fss);
204 else if (s->method == AV_ME_METHOD_NTSS)
205 SEARCH_MV(ntss);
206 else if (s->method == AV_ME_METHOD_TDLS)
207 SEARCH_MV(tdls);
208 else if (s->method == AV_ME_METHOD_TSS)
209 SEARCH_MV(tss);
210 else if (s->method == AV_ME_METHOD_HEXBS)
211 SEARCH_MV(hexbs);
212 else if (s->method == AV_ME_METHOD_UMH) {
213 for (mb_y = 0; mb_y < s->b_height; mb_y++)
214 for (mb_x = 0; mb_x < s->b_width; mb_x++) {
215 const int mb_i = mb_x + mb_y * s->b_width;
216 const int x_mb = mb_x << s->log2_mb_size;
217 const int y_mb = mb_y << s->log2_mb_size;
218 int mv[2] = {x_mb, y_mb};
219
220 AVMotionEstPredictor *preds = me_ctx->preds;
221 preds[0].nb = 0;
222
223 ADD_PRED(preds[0], 0, 0);
224
225 //left mb in current frame
226 if (mb_x > 0)
227 ADD_PRED(preds[0], s->mv_table[0][mb_i - 1][dir][0], s->mv_table[0][mb_i - 1][dir][1]);
228
229 if (mb_y > 0) {
230 //top mb in current frame
231 ADD_PRED(preds[0], s->mv_table[0][mb_i - s->b_width][dir][0], s->mv_table[0][mb_i - s->b_width][dir][1]);
232
233 //top-right mb in current frame
234 if (mb_x + 1 < s->b_width)
235 ADD_PRED(preds[0], s->mv_table[0][mb_i - s->b_width + 1][dir][0], s->mv_table[0][mb_i - s->b_width + 1][dir][1]);
236 //top-left mb in current frame
237 else if (mb_x > 0)
238 ADD_PRED(preds[0], s->mv_table[0][mb_i - s->b_width - 1][dir][0], s->mv_table[0][mb_i - s->b_width - 1][dir][1]);
239 }
240
241 //median predictor
242 if (preds[0].nb == 4) {
243 me_ctx->pred_x = mid_pred(preds[0].mvs[1][0], preds[0].mvs[2][0], preds[0].mvs[3][0]);
244 me_ctx->pred_y = mid_pred(preds[0].mvs[1][1], preds[0].mvs[2][1], preds[0].mvs[3][1]);
245 } else if (preds[0].nb == 3) {
246 me_ctx->pred_x = mid_pred(0, preds[0].mvs[1][0], preds[0].mvs[2][0]);
247 me_ctx->pred_y = mid_pred(0, preds[0].mvs[1][1], preds[0].mvs[2][1]);
248 } else if (preds[0].nb == 2) {
249 me_ctx->pred_x = preds[0].mvs[1][0];
250 me_ctx->pred_y = preds[0].mvs[1][1];
251 } else {
252 me_ctx->pred_x = 0;
253 me_ctx->pred_y = 0;
254 }
255
256 ff_me_search_umh(me_ctx, x_mb, y_mb, mv);
257
258 s->mv_table[0][mb_i][dir][0] = mv[0] - x_mb;
259 s->mv_table[0][mb_i][dir][1] = mv[1] - y_mb;
260 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, me_ctx->mb_size, x_mb, y_mb, mv[0], mv[1], dir);
261 }
262
263 } else if (s->method == AV_ME_METHOD_EPZS) {
264
265 for (mb_y = 0; mb_y < s->b_height; mb_y++)
266 for (mb_x = 0; mb_x < s->b_width; mb_x++) {
267 const int mb_i = mb_x + mb_y * s->b_width;
268 const int x_mb = mb_x << s->log2_mb_size;
269 const int y_mb = mb_y << s->log2_mb_size;
270 int mv[2] = {x_mb, y_mb};
271
272 AVMotionEstPredictor *preds = me_ctx->preds;
273 preds[0].nb = 0;
274 preds[1].nb = 0;
275
276 ADD_PRED(preds[0], 0, 0);
277
278 //left mb in current frame
279 if (mb_x > 0)
280 ADD_PRED(preds[0], s->mv_table[0][mb_i - 1][dir][0], s->mv_table[0][mb_i - 1][dir][1]);
281
282 //top mb in current frame
283 if (mb_y > 0)
284 ADD_PRED(preds[0], s->mv_table[0][mb_i - s->b_width][dir][0], s->mv_table[0][mb_i - s->b_width][dir][1]);
285
286 //top-right mb in current frame
287 if (mb_y > 0 && mb_x + 1 < s->b_width)
288 ADD_PRED(preds[0], s->mv_table[0][mb_i - s->b_width + 1][dir][0], s->mv_table[0][mb_i - s->b_width + 1][dir][1]);
289
290 //median predictor
291 if (preds[0].nb == 4) {
292 me_ctx->pred_x = mid_pred(preds[0].mvs[1][0], preds[0].mvs[2][0], preds[0].mvs[3][0]);
293 me_ctx->pred_y = mid_pred(preds[0].mvs[1][1], preds[0].mvs[2][1], preds[0].mvs[3][1]);
294 } else if (preds[0].nb == 3) {
295 me_ctx->pred_x = mid_pred(0, preds[0].mvs[1][0], preds[0].mvs[2][0]);
296 me_ctx->pred_y = mid_pred(0, preds[0].mvs[1][1], preds[0].mvs[2][1]);
297 } else if (preds[0].nb == 2) {
298 me_ctx->pred_x = preds[0].mvs[1][0];
299 me_ctx->pred_y = preds[0].mvs[1][1];
300 } else {
301 me_ctx->pred_x = 0;
302 me_ctx->pred_y = 0;
303 }
304
305 //collocated mb in prev frame
306 ADD_PRED(preds[0], s->mv_table[1][mb_i][dir][0], s->mv_table[1][mb_i][dir][1]);
307
308 //accelerator motion vector of collocated block in prev frame
309 ADD_PRED(preds[1], s->mv_table[1][mb_i][dir][0] + (s->mv_table[1][mb_i][dir][0] - s->mv_table[2][mb_i][dir][0]),
310 s->mv_table[1][mb_i][dir][1] + (s->mv_table[1][mb_i][dir][1] - s->mv_table[2][mb_i][dir][1]));
311
312 //left mb in prev frame
313 if (mb_x > 0)
314 ADD_PRED(preds[1], s->mv_table[1][mb_i - 1][dir][0], s->mv_table[1][mb_i - 1][dir][1]);
315
316 //top mb in prev frame
317 if (mb_y > 0)
318 ADD_PRED(preds[1], s->mv_table[1][mb_i - s->b_width][dir][0], s->mv_table[1][mb_i - s->b_width][dir][1]);
319
320 //right mb in prev frame
321 if (mb_x + 1 < s->b_width)
322 ADD_PRED(preds[1], s->mv_table[1][mb_i + 1][dir][0], s->mv_table[1][mb_i + 1][dir][1]);
323
324 //bottom mb in prev frame
325 if (mb_y + 1 < s->b_height)
326 ADD_PRED(preds[1], s->mv_table[1][mb_i + s->b_width][dir][0], s->mv_table[1][mb_i + s->b_width][dir][1]);
327
328 ff_me_search_epzs(me_ctx, x_mb, y_mb, mv);
329
330 s->mv_table[0][mb_i][dir][0] = mv[0] - x_mb;
331 s->mv_table[0][mb_i][dir][1] = mv[1] - y_mb;
332 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, s->mb_size, x_mb, y_mb, mv[0], mv[1], dir);
333 }
334 }
335 }
336
337 return ff_filter_frame(ctx->outputs[0], out);
338 }
339
uninit(AVFilterContext * ctx)340 static av_cold void uninit(AVFilterContext *ctx)
341 {
342 MEContext *s = ctx->priv;
343 int i;
344
345 av_frame_free(&s->prev);
346 av_frame_free(&s->cur);
347 av_frame_free(&s->next);
348
349 for (i = 0; i < 3; i++)
350 av_freep(&s->mv_table[i]);
351 }
352
353 static const AVFilterPad mestimate_inputs[] = {
354 {
355 .name = "default",
356 .type = AVMEDIA_TYPE_VIDEO,
357 .filter_frame = filter_frame,
358 .config_props = config_input,
359 },
360 { NULL }
361 };
362
363 static const AVFilterPad mestimate_outputs[] = {
364 {
365 .name = "default",
366 .type = AVMEDIA_TYPE_VIDEO,
367 },
368 { NULL }
369 };
370
371 AVFilter ff_vf_mestimate = {
372 .name = "mestimate",
373 .description = NULL_IF_CONFIG_SMALL("Generate motion vectors."),
374 .priv_size = sizeof(MEContext),
375 .priv_class = &mestimate_class,
376 .uninit = uninit,
377 .query_formats = query_formats,
378 .inputs = mestimate_inputs,
379 .outputs = mestimate_outputs,
380 };
381