avfilter: add vmafmotion filter

Signed-off-by: Ashish Singh <ashk43712@gmail.com>
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>

Changelog

@@ -51,6 +51,7 @@ version <next>:
 - CUDA thumbnail filter
 - V4L2 mem2mem HW assisted codecs
 - Rockchip MPP hardware decoding
+- vmafmotion video filter
 
 
 version 3.3:

doc/filters.texi

@@ -15570,6 +15570,20 @@ vignette='PI/4+random(1)*PI/50':eval=frame
 
 @end itemize
 
+@section vmafmotion
+
+Obtain the average vmaf motion score of a video.
+It is one of the component filters of VMAF.
+
+The obtained average motion score is printed through the logging system.
+
+In the below example the input file @file{ref.mpg} is being processed and score
+is computed.
+
+@example
+ffmpeg -i ref.mpg -lavfi vmafmotion -f null -
+@end example
+
 @section vstack
 Stack input videos vertically.
 

libavfilter/Makefile

@@ -331,6 +331,7 @@ OBJS-$(CONFIG_VFLIP_FILTER)                  += vf_vflip.o
 OBJS-$(CONFIG_VIDSTABDETECT_FILTER)          += vidstabutils.o vf_vidstabdetect.o
 OBJS-$(CONFIG_VIDSTABTRANSFORM_FILTER)       += vidstabutils.o vf_vidstabtransform.o
 OBJS-$(CONFIG_VIGNETTE_FILTER)               += vf_vignette.o
+OBJS-$(CONFIG_VMAFMOTION_FILTER)             += vf_vmafmotion.o framesync.o
 OBJS-$(CONFIG_VSTACK_FILTER)                 += vf_stack.o framesync.o
 OBJS-$(CONFIG_W3FDIF_FILTER)                 += vf_w3fdif.o
 OBJS-$(CONFIG_WAVEFORM_FILTER)               += vf_waveform.o

libavfilter/allfilters.c

@@ -342,6 +342,7 @@ static void register_all(void)
     REGISTER_FILTER(VIDSTABDETECT,  vidstabdetect,  vf);
     REGISTER_FILTER(VIDSTABTRANSFORM, vidstabtransform, vf);
     REGISTER_FILTER(VIGNETTE,       vignette,       vf);
+    REGISTER_FILTER(VMAFMOTION,     vmafmotion,     vf);
     REGISTER_FILTER(VSTACK,         vstack,         vf);
     REGISTER_FILTER(W3FDIF,         w3fdif,         vf);
     REGISTER_FILTER(WAVEFORM,       waveform,       vf);

libavfilter/vf_vmafmotion.c

@@ -0,0 +1,365 @@
+/*
+ * Copyright (c) 2017 Ronald S. Bultje <rsbultje@gmail.com>
+ * Copyright (c) 2017 Ashish Pratap Singh <ashk43712@gmail.com>
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * Calculate VMAF Motion score.
+ */
+
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+#include "avfilter.h"
+#include "drawutils.h"
+#include "formats.h"
+#include "internal.h"
+#include "vmaf_motion.h"
+
+#define BIT_SHIFT 15
+
+static const float FILTER_5[5] = {
+    0.054488685,
+    0.244201342,
+    0.402619947,
+    0.244201342,
+    0.054488685
+};
+
+typedef struct VMAFMotionContext {
+    const AVClass *class;
+    VMAFMotionData data;
+    FILE *stats_file;
+    char *stats_file_str;
+} VMAFMotionContext;
+
+#define OFFSET(x) offsetof(VMAFMotionContext, x)
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
+
+static const AVOption vmafmotion_options[] = {
+    {"stats_file", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(vmafmotion);
+
+static uint64_t image_sad(const uint16_t *img1, const uint16_t *img2, int w,
+                          int h, ptrdiff_t _img1_stride, ptrdiff_t _img2_stride)
+{
+    ptrdiff_t img1_stride = _img1_stride / sizeof(*img1);
+    ptrdiff_t img2_stride = _img2_stride / sizeof(*img2);
+    uint64_t sum = 0;
+    int i, j;
+
+    for (i = 0; i < h; i++) {
+        for (j = 0; j < w; j++) {
+            sum += abs(img1[j] - img2[j]);
+        }
+        img1 += img1_stride;
+        img2 += img2_stride;
+    }
+
+    return sum;
+}
+
+static void convolution_x(const uint16_t *filter, int filt_w, const uint16_t *src,
+                          uint16_t *dst, int w, int h, ptrdiff_t _src_stride,
+                          ptrdiff_t _dst_stride)
+{
+    ptrdiff_t src_stride = _src_stride / sizeof(*src);
+    ptrdiff_t dst_stride = _dst_stride / sizeof(*dst);
+    int radius = filt_w / 2;
+    int borders_left = radius;
+    int borders_right = w - (filt_w - radius);
+    int i, j, k;
+    int sum = 0;
+
+    for (i = 0; i < h; i++) {
+        for (j = 0; j < borders_left; j++) {
+            sum = 0;
+            for (k = 0; k < filt_w; k++) {
+                int j_tap = FFABS(j - radius + k);
+                if (j_tap >= w) {
+                    j_tap = w - (j_tap - w + 1);
+                }
+                sum += filter[k] * src[i * src_stride + j_tap];
+            }
+            dst[i * dst_stride + j] = sum >> BIT_SHIFT;
+        }
+
+        for (j = borders_left; j < borders_right; j++) {
+            int sum = 0;
+            for (k = 0; k < filt_w; k++) {
+                sum += filter[k] * src[i * src_stride + j - radius + k];
+            }
+            dst[i * dst_stride + j] = sum >> BIT_SHIFT;
+        }
+
+        for (j = borders_right; j < w; j++) {
+            sum = 0;
+            for (k = 0; k < filt_w; k++) {
+                int j_tap = FFABS(j - radius + k);
+                if (j_tap >= w) {
+                    j_tap = w - (j_tap - w + 1);
+                }
+                sum += filter[k] * src[i * src_stride + j_tap];
+            }
+            dst[i * dst_stride + j] = sum >> BIT_SHIFT;
+        }
+    }
+}
+
+#define conv_y_fn(type, bits) \
+static void convolution_y_##bits##bit(const uint16_t *filter, int filt_w, \
+                                      const uint8_t *_src, uint16_t *dst, \
+                                      int w, int h, ptrdiff_t _src_stride, \
+                                      ptrdiff_t _dst_stride) \
+{ \
+    const type *src = (const type *) _src; \
+    ptrdiff_t src_stride = _src_stride / sizeof(*src); \
+    ptrdiff_t dst_stride = _dst_stride / sizeof(*dst); \
+    int radius = filt_w / 2; \
+    int borders_top = radius; \
+    int borders_bottom = h - (filt_w - radius); \
+    int i, j, k; \
+    int sum = 0; \
+    \
+    for (i = 0; i < borders_top; i++) { \
+        for (j = 0; j < w; j++) { \
+            sum = 0; \
+            for (k = 0; k < filt_w; k++) { \
+                int i_tap = FFABS(i - radius + k); \
+                if (i_tap >= h) { \
+                    i_tap = h - (i_tap - h + 1); \
+                } \
+                sum += filter[k] * src[i_tap * src_stride + j]; \
+            } \
+            dst[i * dst_stride + j] = sum >> bits; \
+        } \
+    } \
+    for (i = borders_top; i < borders_bottom; i++) { \
+        for (j = 0; j < w; j++) { \
+            sum = 0; \
+            for (k = 0; k < filt_w; k++) { \
+                sum += filter[k] * src[(i - radius + k) * src_stride + j]; \
+            } \
+            dst[i * dst_stride + j] = sum >> bits; \
+        } \
+    } \
+    for (i = borders_bottom; i < h; i++) { \
+        for (j = 0; j < w; j++) { \
+            sum = 0; \
+            for (k = 0; k < filt_w; k++) { \
+                int i_tap = FFABS(i - radius + k); \
+                if (i_tap >= h) { \
+                    i_tap = h - (i_tap - h + 1); \
+                } \
+                sum += filter[k] * src[i_tap * src_stride + j]; \
+            } \
+            dst[i * dst_stride + j] = sum >> bits; \
+        } \
+    } \
+}
+
+conv_y_fn(uint8_t, 8);
+conv_y_fn(uint16_t, 10);
+
+static void vmafmotiondsp_init(VMAFMotionDSPContext *dsp, int bpp) {
+    dsp->convolution_x = convolution_x;
+    dsp->convolution_y = bpp == 10 ? convolution_y_10bit : convolution_y_8bit;
+    dsp->sad = image_sad;
+}
+
+double ff_vmafmotion_process(VMAFMotionData *s, AVFrame *ref)
+{
+    double score;
+
+    s->vmafdsp.convolution_y(s->filter, 5, ref->data[0], s->temp_data,
+                             s->width, s->height, ref->linesize[0], s->stride);
+    s->vmafdsp.convolution_x(s->filter, 5, s->temp_data, s->blur_data[0],
+                             s->width, s->height, s->stride, s->stride);
+
+    if (!s->nb_frames) {
+        score = 0.0;
+    } else {
+        uint64_t sad = s->vmafdsp.sad(s->blur_data[1], s->blur_data[0],
+                                      s->width, s->height, s->stride, s->stride);
+        // the output score is always normalized to 8 bits
+        score = (double) (sad * 1.0 / (s->width * s->height << (BIT_SHIFT - 8)));
+    }
+
+    FFSWAP(uint16_t *, s->blur_data[0], s->blur_data[1]);
+    s->nb_frames++;
+    s->motion_sum += score;
+
+    return score;
+}
+
+static void set_meta(AVDictionary **metadata, const char *key, float d)
+{
+    char value[128];
+    snprintf(value, sizeof(value), "%0.2f", d);
+    av_dict_set(metadata, key, value, 0);
+}
+
+static void do_vmafmotion(AVFilterContext *ctx, AVFrame *ref)
+{
+    VMAFMotionContext *s = ctx->priv;
+    double score;
+
+    score = ff_vmafmotion_process(&s->data, ref);
+    set_meta(&ref->metadata, "lavfi.vmafmotion.score", score);
+    if (s->stats_file) {
+        fprintf(s->stats_file,
+                "n:%"PRId64" motion:%0.2lf\n", s->data.nb_frames, score);
+    }
+}
+
+
+int ff_vmafmotion_init(VMAFMotionData *s,
+                       int w, int h, enum AVPixelFormat fmt)
+{
+    size_t data_sz;
+    int i;
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
+
+    s->width = w;
+    s->height = h;
+    s->stride = FFALIGN(w * sizeof(uint16_t), 32);
+
+    data_sz = (size_t) s->stride * h;
+    if (!(s->blur_data[0] = av_malloc(data_sz)) ||
+        !(s->blur_data[1] = av_malloc(data_sz)) ||
+        !(s->temp_data    = av_malloc(data_sz))) {
+        return AVERROR(ENOMEM);
+    }
+
+    for (i = 0; i < 5; i++) {
+        s->filter[i] = lrint(FILTER_5[i] * (1 << BIT_SHIFT));
+    }
+
+    vmafmotiondsp_init(&s->vmafdsp, desc->comp[0].depth);
+
+    return 0;
+}
+
+static int query_formats(AVFilterContext *ctx)
+{
+    static const enum AVPixelFormat pix_fmts[] = {
+        AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV420P,
+        AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV420P10,
+        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_ref(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx  = inlink->dst;
+    VMAFMotionContext *s = ctx->priv;
+
+    return ff_vmafmotion_init(&s->data, ctx->inputs[0]->w,
+                              ctx->inputs[0]->h, ctx->inputs[0]->format);
+}
+
+double ff_vmafmotion_uninit(VMAFMotionData *s)
+{
+    av_free(s->blur_data[0]);
+    av_free(s->blur_data[1]);
+    av_free(s->temp_data);
+
+    return s->nb_frames > 0 ? s->motion_sum / s->nb_frames : 0.0;
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *ref)
+{
+    AVFilterContext *ctx = inlink->dst;
+    do_vmafmotion(ctx, ref);
+    return ff_filter_frame(ctx->outputs[0], ref);
+}
+
+static av_cold int init(AVFilterContext *ctx)
+{
+    VMAFMotionContext *s = ctx->priv;
+
+    if (s->stats_file_str) {
+        if (!strcmp(s->stats_file_str, "-")) {
+            s->stats_file = stdout;
+        } else {
+            s->stats_file = fopen(s->stats_file_str, "w");
+            if (!s->stats_file) {
+                int err = AVERROR(errno);
+                char buf[128];
+                av_strerror(err, buf, sizeof(buf));
+                av_log(ctx, AV_LOG_ERROR, "Could not open stats file %s: %s\n",
+                       s->stats_file_str, buf);
+                return err;
+            }
+        }
+    }
+
+    return 0;
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    VMAFMotionContext *s = ctx->priv;
+    double avg_motion = ff_vmafmotion_uninit(&s->data);
+
+    if (s->data.nb_frames > 0) {
+        av_log(ctx, AV_LOG_INFO, "VMAF Motion avg: %.3f\n", avg_motion);
+    }
+
+    if (s->stats_file && s->stats_file != stdout)
+        fclose(s->stats_file);
+}
+
+static const AVFilterPad vmafmotion_inputs[] = {
+    {
+        .name         = "reference",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = filter_frame,
+        .config_props = config_input_ref,
+    },
+    { NULL }
+};
+
+static const AVFilterPad vmafmotion_outputs[] = {
+    {
+        .name          = "default",
+        .type          = AVMEDIA_TYPE_VIDEO,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_vmafmotion = {
+    .name          = "vmafmotion",
+    .description   = NULL_IF_CONFIG_SMALL("Calculate the VMAF Motion score."),
+    .init          = init,
+    .uninit        = uninit,
+    .query_formats = query_formats,
+    .priv_size     = sizeof(VMAFMotionContext),
+    .priv_class    = &vmafmotion_class,
+    .inputs        = vmafmotion_inputs,
+    .outputs       = vmafmotion_outputs,
+};

libavfilter/vmaf_motion.h

@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2017 Ronald S. Bultje <rsbultje@gmail.com>
+ * Copyright (c) 2017 Ashish Pratap Singh <ashk43712@gmail.com>
+ *
+ * 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
+ */
+
+#ifndef AVFILTER_VMAFMOTION_H
+#define AVFILTER_VMAFMOTION_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include "video.h"
+
+typedef struct VMAFMotionDSPContext {
+    uint64_t (*sad)(const uint16_t *img1, const uint16_t *img2, int w, int h,
+                    ptrdiff_t img1_stride, ptrdiff_t img2_stride);
+    void (*convolution_x)(const uint16_t *filter, int filt_w, const uint16_t *src,
+                          uint16_t *dst, int w, int h, ptrdiff_t src_stride,
+                          ptrdiff_t dst_stride);
+    void (*convolution_y)(const uint16_t *filter, int filt_w, const uint8_t *src,
+                          uint16_t *dst, int w, int h, ptrdiff_t src_stride,
+                          ptrdiff_t dst_stride);
+} VMAFMotionDSPContext;
+
+void ff_vmafmotion_init_x86(VMAFMotionDSPContext *dsp);
+
+typedef struct VMAFMotionData {
+    uint16_t filter[5];
+    int width;
+    int height;
+    ptrdiff_t stride;
+    uint16_t *blur_data[2 /* cur, prev */];
+    uint16_t *temp_data;
+    double motion_sum;
+    uint64_t nb_frames;
+    VMAFMotionDSPContext vmafdsp;
+} VMAFMotionData;
+
+int ff_vmafmotion_init(VMAFMotionData *data, int w, int h, enum AVPixelFormat fmt);
+double ff_vmafmotion_process(VMAFMotionData *data, AVFrame *frame);
+double ff_vmafmotion_uninit(VMAFMotionData *data);
+
+#endif /* AVFILTER_VMAFMOTION_H */