2 * Copyright (c) 2016 Davinder Singh (DSM_) <ds.mudhar<@gmail.com>
4 * This file is part of FFmpeg.
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.
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.
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
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"
34 typedef struct MEContext {
36 AVMotionEstContext me_ctx;
37 int method; ///< motion estimation method
39 int mb_size; ///< macroblock size
40 int search_param; ///< search parameter
41 int b_width, b_height, b_count;
44 AVFrame *prev, *cur, *next;
46 int (*mv_table[3])[2][2]; ///< motion vectors of current & prev 2 frames
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 }
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 },
69 AVFILTER_DEFINE_CLASS(mestimate);
71 static int query_formats(AVFilterContext *ctx)
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,
80 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
85 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
87 return AVERROR(ENOMEM);
88 return ff_set_common_formats(ctx, fmts_list);
91 static int config_input(AVFilterLink *inlink)
93 MEContext *s = inlink->dst->priv;
96 s->log2_mb_size = av_ceil_log2_c(s->mb_size);
97 s->mb_size = 1 << s->log2_mb_size;
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;
103 for (i = 0; i < 3; i++) {
104 s->mv_table[i] = av_mallocz_array(s->b_count, sizeof(*s->mv_table[0]));
106 return AVERROR(ENOMEM);
109 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);
114 static void add_mv_data(AVMotionVector *mv, int mb_size,
115 int x, int y, int x_mv, int y_mv, int dir)
119 mv->dst_x = x + (mb_size >> 1);
120 mv->dst_y = y + (mb_size >> 1);
121 mv->src_x = x_mv + (mb_size >> 1);
122 mv->src_y = y_mv + (mb_size >> 1);
123 mv->source = dir ? 1 : -1;
127 #define SEARCH_MV(method)\
129 for (mb_y = 0; mb_y < s->b_height; mb_y++)\
130 for (mb_x = 0; mb_x < s->b_width; mb_x++) {\
131 const int x_mb = mb_x << s->log2_mb_size;\
132 const int y_mb = mb_y << s->log2_mb_size;\
133 int mv[2] = {x_mb, y_mb};\
134 ff_me_search_##method(me_ctx, x_mb, y_mb, mv);\
135 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, me_ctx->mb_size, x_mb, y_mb, mv[0], mv[1], dir);\
139 #define ADD_PRED(preds, px, py)\
141 preds.mvs[preds.nb][0] = px;\
142 preds.mvs[preds.nb][1] = py;\
146 static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
148 AVFilterContext *ctx = inlink->dst;
149 MEContext *s = ctx->priv;
150 AVMotionEstContext *me_ctx = &s->me_ctx;
154 int32_t mv_count = 0;
157 if (frame->pts == AV_NOPTS_VALUE) {
158 ret = ff_filter_frame(ctx->outputs[0], frame);
162 av_frame_free(&s->prev);
167 s->mv_table[2] = memcpy(s->mv_table[2], s->mv_table[1], sizeof(*s->mv_table[1]) * s->b_count);
168 s->mv_table[1] = memcpy(s->mv_table[1], s->mv_table[0], sizeof(*s->mv_table[0]) * s->b_count);
171 s->cur = av_frame_clone(frame);
173 return AVERROR(ENOMEM);
179 out = av_frame_clone(s->cur);
181 return AVERROR(ENOMEM);
183 sd = av_frame_new_side_data(out, AV_FRAME_DATA_MOTION_VECTORS, 2 * s->b_count * sizeof(AVMotionVector));
186 return AVERROR(ENOMEM);
189 me_ctx->data_cur = s->cur->data[0];
190 me_ctx->linesize = s->cur->linesize[0];
192 for (dir = 0; dir < 2; dir++) {
193 me_ctx->data_ref = (dir ? s->next : s->prev)->data[0];
195 if (s->method == AV_ME_METHOD_DS)
197 else if (s->method == AV_ME_METHOD_ESA)
199 else if (s->method == AV_ME_METHOD_FSS)
201 else if (s->method == AV_ME_METHOD_NTSS)
203 else if (s->method == AV_ME_METHOD_TDLS)
205 else if (s->method == AV_ME_METHOD_TSS)
207 else if (s->method == AV_ME_METHOD_HEXBS)
209 else if (s->method == AV_ME_METHOD_UMH) {
210 for (mb_y = 0; mb_y < s->b_height; mb_y++)
211 for (mb_x = 0; mb_x < s->b_width; mb_x++) {
212 const int mb_i = mb_x + mb_y * s->b_width;
213 const int x_mb = mb_x << s->log2_mb_size;
214 const int y_mb = mb_y << s->log2_mb_size;
215 int mv[2] = {x_mb, y_mb};
217 AVMotionEstPredictor *preds = me_ctx->preds;
220 ADD_PRED(preds[0], 0, 0);
222 //left mb in current frame
224 ADD_PRED(preds[0], s->mv_table[0][mb_i - 1][dir][0], s->mv_table[0][mb_i - 1][dir][1]);
227 //top mb in current frame
228 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]);
230 //top-right mb in current frame
231 if (mb_x + 1 < s->b_width)
232 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]);
233 //top-left mb in current frame
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]);
239 if (preds[0].nb == 4) {
240 me_ctx->pred_x = mid_pred(preds[0].mvs[1][0], preds[0].mvs[2][0], preds[0].mvs[3][0]);
241 me_ctx->pred_y = mid_pred(preds[0].mvs[1][1], preds[0].mvs[2][1], preds[0].mvs[3][1]);
242 } else if (preds[0].nb == 3) {
243 me_ctx->pred_x = mid_pred(0, preds[0].mvs[1][0], preds[0].mvs[2][0]);
244 me_ctx->pred_y = mid_pred(0, preds[0].mvs[1][1], preds[0].mvs[2][1]);
245 } else if (preds[0].nb == 2) {
246 me_ctx->pred_x = preds[0].mvs[1][0];
247 me_ctx->pred_y = preds[0].mvs[1][1];
253 ff_me_search_umh(me_ctx, x_mb, y_mb, mv);
255 s->mv_table[0][mb_i][dir][0] = mv[0] - x_mb;
256 s->mv_table[0][mb_i][dir][1] = mv[1] - y_mb;
257 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, me_ctx->mb_size, x_mb, y_mb, mv[0], mv[1], dir);
260 } else if (s->method == AV_ME_METHOD_EPZS) {
262 for (mb_y = 0; mb_y < s->b_height; mb_y++)
263 for (mb_x = 0; mb_x < s->b_width; mb_x++) {
264 const int mb_i = mb_x + mb_y * s->b_width;
265 const int x_mb = mb_x << s->log2_mb_size;
266 const int y_mb = mb_y << s->log2_mb_size;
267 int mv[2] = {x_mb, y_mb};
269 AVMotionEstPredictor *preds = me_ctx->preds;
273 ADD_PRED(preds[0], 0, 0);
275 //left mb in current frame
277 ADD_PRED(preds[0], s->mv_table[0][mb_i - 1][dir][0], s->mv_table[0][mb_i - 1][dir][1]);
279 //top mb in current frame
281 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]);
283 //top-right mb in current frame
284 if (mb_y > 0 && mb_x + 1 < s->b_width)
285 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]);
288 if (preds[0].nb == 4) {
289 me_ctx->pred_x = mid_pred(preds[0].mvs[1][0], preds[0].mvs[2][0], preds[0].mvs[3][0]);
290 me_ctx->pred_y = mid_pred(preds[0].mvs[1][1], preds[0].mvs[2][1], preds[0].mvs[3][1]);
291 } else if (preds[0].nb == 3) {
292 me_ctx->pred_x = mid_pred(0, preds[0].mvs[1][0], preds[0].mvs[2][0]);
293 me_ctx->pred_y = mid_pred(0, preds[0].mvs[1][1], preds[0].mvs[2][1]);
294 } else if (preds[0].nb == 2) {
295 me_ctx->pred_x = preds[0].mvs[1][0];
296 me_ctx->pred_y = preds[0].mvs[1][1];
302 //collocated mb in prev frame
303 ADD_PRED(preds[0], s->mv_table[1][mb_i][dir][0], s->mv_table[1][mb_i][dir][1]);
305 //accelerator motion vector of collocated block in prev frame
306 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]),
307 s->mv_table[1][mb_i][dir][1] + (s->mv_table[1][mb_i][dir][1] - s->mv_table[2][mb_i][dir][1]));
309 //left mb in prev frame
311 ADD_PRED(preds[1], s->mv_table[1][mb_i - 1][dir][0], s->mv_table[1][mb_i - 1][dir][1]);
313 //top mb in prev frame
315 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]);
317 //right mb in prev frame
318 if (mb_x + 1 < s->b_width)
319 ADD_PRED(preds[1], s->mv_table[1][mb_i + 1][dir][0], s->mv_table[1][mb_i + 1][dir][1]);
321 //bottom mb in prev frame
322 if (mb_y + 1 < s->b_height)
323 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]);
325 ff_me_search_epzs(me_ctx, x_mb, y_mb, mv);
327 s->mv_table[0][mb_i][dir][0] = mv[0] - x_mb;
328 s->mv_table[0][mb_i][dir][1] = mv[1] - y_mb;
329 add_mv_data(((AVMotionVector *) sd->data) + mv_count++, s->mb_size, x_mb, y_mb, mv[0], mv[1], dir);
334 return ff_filter_frame(ctx->outputs[0], out);
337 static av_cold void uninit(AVFilterContext *ctx)
339 MEContext *s = ctx->priv;
342 av_frame_free(&s->prev);
343 av_frame_free(&s->cur);
344 av_frame_free(&s->next);
346 for (i = 0; i < 3; i++)
347 av_freep(&s->mv_table[i]);
350 static const AVFilterPad mestimate_inputs[] = {
353 .type = AVMEDIA_TYPE_VIDEO,
354 .filter_frame = filter_frame,
355 .config_props = config_input,
360 static const AVFilterPad mestimate_outputs[] = {
363 .type = AVMEDIA_TYPE_VIDEO,
368 const AVFilter ff_vf_mestimate = {
370 .description = NULL_IF_CONFIG_SMALL("Generate motion vectors."),
371 .priv_size = sizeof(MEContext),
372 .priv_class = &mestimate_class,
374 .query_formats = query_formats,
375 .inputs = mestimate_inputs,
376 .outputs = mestimate_outputs,