[FFmpeg-devel] [Patch][GSoC] Motion Estimation filter

Paul B Mahol onemda at gmail.com
Sat Apr 9 21:17:25 CEST 2016


On 4/9/16, Davinder Singh <ds.mudhar at gmail.com> wrote:
> hi
> here's the new patch.
>
> changes: fixed bugs, added documentation
> added a step option, by default, while searching for macro-block, it moves
> by 1 px in search area. this can be increased now.
> uses absolute values instead of squares.
> vector starts from center of macro-block, before it was at top-left.
>
> There were a lot of "Past duration 0.999992 too large" warnings each frame
> while encoding, perhaps its because the frame lags by 1 (as I store the
> requested frame in MEContext->next)
> this is fixed by changing: out_frame->pts = MEContext->next->pts
>
> details of the filter:
> The current filter is most basic block estimation technique which does full
> search, which makes it quite slow, speed depends on the search area in
> which the block is searched and it is independent of the block size. if we
> take N x N block size, and video is W x H. Then total no of blocks will be
> WxH / N^2. if R is search parameter then the no of places block is
> searched: (2R + 1)^2, and for each block there will be N^2 iterations (for
> comparing each pixel value which gives MAD. the one with minimum MAD is
> used to get MV), so for all blocks in a frame, there will be WxH/N^2 *
> (2R+1)^2*N^2 iterations = WxH * (2R+1)^2 iterations.
> e.g. for a 720p video, R = 7 it will be 1280*720*15^2 = 207360000 iteration
> per frame.
>
> The current implementation only use first plane (luma) for estimation.
>
> Obviously I have to implement a faster algorithm, I chose this one for
> simplicity.

> From 01f21e83d92389355105d4c9ba0ac1b835e343cb Mon Sep 17 00:00:00 2001
> From: dsmudhar <ds.mudhar at gmail.com>
> Date: Sun, 10 Apr 2016 00:01:23 +0530
> Subject: [PATCH] added motion estimation filter
>
> ---
>  doc/filters.texi           |  16 +++
>  libavfilter/Makefile       |   1 +
>  libavfilter/allfilters.c   |   1 +
>  libavfilter/vf_mestimate.c | 245 +++++++++++++++++++++++++++++++++++++++++++++
>  4 files changed, 263 insertions(+)
>  create mode 100644 libavfilter/vf_mestimate.c
>
> diff --git a/doc/filters.texi b/doc/filters.texi
> index 82be06d..85757a1 100644
> --- a/doc/filters.texi
> +++ b/doc/filters.texi
> @@ -8933,6 +8933,22 @@ format=rgb24,mergeplanes=0x000102:yuv444p
>  @end example
>  @end itemize
>
> + at section mestimate
> +
> +Estimates the motion and generates motion vectors using block matching algorithm.
> +
> +This filter accepts the following options:
> + at table @option
> + at item block
> +Set macroblock size. Default is @code{16}.
> +
> + at item search
> +Set search parameter. Default is @code{7}.
> +
> + at item step
> +Set step for movement of reference macroblock in search area. Default is @code{1}.
> + at end table
> +
>  @section metadata, ametadata
>
>  Manipulate frame metadata.
> diff --git a/libavfilter/Makefile b/libavfilter/Makefile
> index 3a3de48..72b75d8 100644
> --- a/libavfilter/Makefile
> +++ b/libavfilter/Makefile
> @@ -198,6 +198,7 @@ OBJS-$(CONFIG_LUTYUV_FILTER)                 += vf_lut.o
>  OBJS-$(CONFIG_MASKEDMERGE_FILTER)            += vf_maskedmerge.o framesync.o
>  OBJS-$(CONFIG_MCDEINT_FILTER)                += vf_mcdeint.o
>  OBJS-$(CONFIG_MERGEPLANES_FILTER)            += vf_mergeplanes.o framesync.o
> +OBJS-$(CONFIG_MESTIMATE_FILTER)              += vf_mestimate.o
>  OBJS-$(CONFIG_METADATA_FILTER)               += f_metadata.o
>  OBJS-$(CONFIG_MPDECIMATE_FILTER)             += vf_mpdecimate.o
>  OBJS-$(CONFIG_NEGATE_FILTER)                 += vf_lut.o
> diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
> index b6f4a2c..6e86fd8 100644
> --- a/libavfilter/allfilters.c
> +++ b/libavfilter/allfilters.c
> @@ -219,6 +219,7 @@ void avfilter_register_all(void)
>      REGISTER_FILTER(MASKEDMERGE,    maskedmerge,    vf);
>      REGISTER_FILTER(MCDEINT,        mcdeint,        vf);
>      REGISTER_FILTER(MERGEPLANES,    mergeplanes,    vf);
> +    REGISTER_FILTER(MESTIMATE,      mestimate,      vf);
>      REGISTER_FILTER(METADATA,       metadata,       vf);
>      REGISTER_FILTER(MPDECIMATE,     mpdecimate,     vf);
>      REGISTER_FILTER(NEGATE,         negate,         vf);
> diff --git a/libavfilter/vf_mestimate.c b/libavfilter/vf_mestimate.c
> new file mode 100644
> index 0000000..448223c
> --- /dev/null
> +++ b/libavfilter/vf_mestimate.c
> @@ -0,0 +1,245 @@
> +/**
> + *
> + * Developed by Davinder Singh (DSM_ / @dsmudhar) during GSoC 2016
> + * As qualification task: Basic but working motion estimation filter
> + *
> + * Used vf_w3fdif.c as base, needed two frames for bi-directional prediction.
> + *
> + * The filter uses block matching exhaustive search algorithm
> + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +#include "libavutil/common.h"
> +#include "libavutil/imgutils.h"
> +#include "libavutil/opt.h"
> +#include "libavutil/pixdesc.h"
> +#include "libavutil/motion_vector.h"
> +#include "avfilter.h"
> +#include "formats.h"
> +#include "internal.h"
> +#include "video.h"
> +
> +typedef struct MEContext {
> +    const AVClass *class;
> +    AVMotionVector *mvs; ///< motion vectors
> +    AVFrame *prev, *cur, *next;  ///< previous, current, next frames
> +    int block_size; ///< block size
> +    int reg_size; ///< search parameter
> +    int32_t mv_count; ///< no of motion vectors per frame
> +    int step; ///< step for movement of reference block in search area
> +
> +} MEContext;
> +
> +#define OFFSET(x) offsetof(MEContext, x)
> +#define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
> +#define CONST(name, help, val, unit) { name, help, 0, AV_OPT_TYPE_CONST, {.i64=val}, 0, 0, FLAGS, unit }
> +
> +static const AVOption mestimate_options[] = {

Nit: You could add alghoritm option here, and only one value to pick
for now: EBMA.

> +    { "block", "specify the macroblock size", OFFSET(block_size), AV_OPT_TYPE_INT, {.i64=16}, 4, INT_MAX, FLAGS, "block" },
> +    { "search",  "specify the search parameter", OFFSET(reg_size), AV_OPT_TYPE_INT, {.i64=7}, 4, INT_MAX, FLAGS, "search" },
> +    { "step",  "specify step for movement of reference block in search area", OFFSET(step), AV_OPT_TYPE_INT, {.i64=1}, 1, INT_MAX, FLAGS, "step" },
> +    { NULL }
> +};
> +
> +AVFILTER_DEFINE_CLASS(mestimate);
> +
> +static int query_formats(AVFilterContext *ctx)
> +{
> +    static const enum AVPixelFormat pix_fmts[] = {
> +        AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
> +        AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
> +        AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
> +        AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
> +        AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
> +        AV_PIX_FMT_YUVJ411P,
> +        AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
> +        AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRAP,
> +        AV_PIX_FMT_GRAY8,
> +        AV_PIX_FMT_NONE
> +    };
> +
> +    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
> +    if (!fmts_list)
> +        return AVERROR(ENOMEM);
> +    return ff_set_common_formats(ctx, fmts_list);
> +}
> +
> +static int config_input(AVFilterLink *inlink)
> +{
> +    MEContext *s = inlink->dst->priv;
> +    int nb_blocks_y = inlink->h / s->block_size;
> +    int nb_blocks_x = inlink->w / s->block_size;
> +
> +    s->mvs = av_malloc_array(nb_blocks_x * nb_blocks_y, 2 * sizeof(AVMotionVector));
> +    if (!s->mvs)
> +        return AVERROR(ENOMEM);
> +
> +    return 0;
> +}
> +
> +static int64_t get_mad(MEContext *s, int x_cur, int y_cur, int x_sb, int y_sb, int direction)
> +{
> +    // dir = 0 => source = -1 => forward prediction  => frame k - 1 as reference
> +    // dir = 1 => source = +1 => backward prediction => frame k + 1 as reference
> +    uint8_t *buf_ref = (direction ? s->next : s->prev)->data[0];
> +    uint8_t *buf_cur = s->cur->data[0];
> +    int stride_ref = (direction ? s->next : s->prev)->linesize[0];
> +    int stride_cur = s->cur->linesize[0];
> +    int64_t mad = 0;
> +    int x, y;
> +
> +    buf_ref += y_sb * stride_ref;
> +    buf_cur += y_cur * stride_cur;
> +
> +    for (y = 0; y < s->block_size; y++)
> +        for (x = 0; x < s->block_size; x++) {
> +            int diff = buf_ref[y * stride_ref + x + x_sb] - buf_cur[y * stride_cur + x + x_cur];
> +            mad += FFABS(diff);
> +        }
> +
> +    return mad;
> +}
> +
> +static void add_mv_data(AVMotionVector *mv, int block_size,
> +                        int dst_x, int dst_y, int src_x, int src_y, int direction)
> +{
> +    mv->w = block_size;
> +    mv->h = block_size;
> +    mv->dst_x = dst_x;
> +    mv->dst_y = dst_y;
> +    mv->src_x = src_x;
> +    mv->src_y = src_y;
> +    mv->source = direction ? 1 : -1;
> +    mv->flags = 0;
> +}
> +
> +static void get_motion_vector(AVFilterLink *inlink, int x_cur, int y_cur, int dir)
> +{
> +    AVFilterContext *ctx = inlink->dst;
> +    MEContext *s = ctx->priv;
> +
> +    int i, j, x_sb, y_sb, dx = 0, dy = 0;
> +    int sign_i = 1, sign_j = 1;
> +    int y_sb_max = av_clip(y_cur + s->reg_size, 0, inlink->h - s->block_size - 1);
> +    int x_sb_max = av_clip(x_cur + s->reg_size, 0, inlink->w - s->block_size - 1);

Add check if image size is lower than block_size... Otherwise this
will crash I guess.

> +    int64_t mad, mad_min = INT64_MAX;

Use unsigned.

> +
> +    for (i = 0; i <= s->reg_size; i = sign_i ? -(i - s->step) : -i, sign_i = !sign_i) {
> +        y_sb = y_cur + i;
> +
> +        if (y_sb < 0 || y_sb > y_sb_max)
> +            continue;
> +
> +        for (j = 0; j <= s->reg_size; j = sign_j ? -(j - s->step) : -j, sign_j = !sign_j) {
> +            x_sb = x_cur + j;
> +
> +            if (x_sb < 0 || x_sb > x_sb_max)
> +                continue;
> +
> +            if ((mad = get_mad(s, x_cur, y_cur, x_sb, y_sb, dir)) < mad_min) {
> +                mad_min = mad;
> +                dx = x_cur - x_sb;
> +                dy = y_cur - y_sb;
> +            }
> +        }
> +    }

All this sign operation seems silly to me, you could just search all
area without those steps that
takes extra instructions?

> +
> +    if (dx || dy) {
> +        x_cur += s->block_size / 2;
> +        y_cur += s->block_size / 2;
> +        add_mv_data(s->mvs + s->mv_count++, s->block_size, x_cur + dx, y_cur + dy, x_cur, y_cur, dir);
> +    }
> +}
> +
> +static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
> +{
> +    AVFilterContext *ctx = inlink->dst;
> +    MEContext *s = ctx->priv;
> +    AVFrameSideData *sd;
> +    int x, y, dir;
> +
> +    av_frame_free(&s->prev);
> +    s->mv_count = 0;
> +    s->prev = s->cur;
> +    s->cur  = s->next;
> +    s->next = frame;
> +
> +    if (s->cur) {
> +        for (y = 0; y < inlink->h; y += s->block_size)
> +            for (x = 0; x < inlink->w; x+= s->block_size)
> +                for (dir = 0; dir < 2; dir++)
> +                    get_motion_vector(inlink, x, y, dir);
> +    } else { // no vectors will be generated if cloned, so skipping for first frame (s->prev, s->next are null)
> +        s->cur = av_frame_clone(s->next);
> +        if (!s->cur)
> +            return AVERROR(ENOMEM);
> +    }
> +
> +    AVFrame *out = av_frame_clone(s->cur);
> +    if (!out)
> +        return AVERROR(ENOMEM);
> +    out->pts = s->next->pts;
> +
> +    if (s->mv_count) {
> +        sd = av_frame_new_side_data(out, AV_FRAME_DATA_MOTION_VECTORS, s->mv_count * sizeof(AVMotionVector));
> +        if (!sd)
> +            return AVERROR(ENOMEM);
> +        memcpy(sd->data, s->mvs, s->mv_count * sizeof(AVMotionVector));
> +    }
> +
> +    return ff_filter_frame(ctx->outputs[0], out);
> +}
> +
> +static av_cold void uninit(AVFilterContext *ctx)
> +{
> +    MEContext *s = ctx->priv;
> +
> +    av_frame_free(&s->prev);
> +    av_frame_free(&s->cur );
> +    av_frame_free(&s->next);
> +    av_freep(&s->mvs);
> +}
> +
> +static const AVFilterPad mestimate_inputs[] = {
> +    {
> +        .name          = "default",
> +        .type          = AVMEDIA_TYPE_VIDEO,
> +        .filter_frame  = filter_frame,
> +        .config_props  = config_input,
> +    },
> +    { NULL }
> +};
> +
> +static const AVFilterPad mestimate_outputs[] = {
> +    {
> +        .name          = "default",
> +        .type          = AVMEDIA_TYPE_VIDEO,
> +    },
> +    { NULL }
> +};
> +
> +AVFilter ff_vf_mestimate = {
> +    .name          = "mestimate",
> +    .description   = NULL_IF_CONFIG_SMALL("Generates motion vectors."),

Generate.

> +    .priv_size     = sizeof(MEContext),
> +    .priv_class    = &mestimate_class,
> +    .uninit        = uninit,
> +    .query_formats = query_formats,
> +    .inputs        = mestimate_inputs,
> +    .outputs       = mestimate_outputs,
> +};
> --
> 2.6.4 (Apple Git-63)
>


More information about the ffmpeg-devel mailing list