[FFmpeg-devel] [PATCH V5] Add a filter implementing HDR image generation from a single exposure using deep CNNs

Li, Zhong zhong.li at intel.com
Thu Nov 15 14:22:18 EET 2018


> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces at ffmpeg.org] On Behalf
> Of Liu Steven
> Sent: Thursday, November 15, 2018 5:40 PM
> To: FFmpeg development discussions and patches
> <ffmpeg-devel at ffmpeg.org>
> Cc: Liu Steven <lq at chinaffmpeg.org>
> Subject: Re: [FFmpeg-devel] [PATCH V5] Add a filter implementing HDR
> image generation from a single exposure using deep CNNs
> 
> 
> 
> > 在 2018年11月14日,下午8:15,Guo, Yejun <yejun.guo at intel.com> 写
> 道:
> >
> > see the algorithm's paper and code below.
> >
> > the filter's parameter looks like:
> >
> sdr2hdr=model_filename=/path_to_tensorflow_graph.pb:out_fmt=gbrp10le
> >
> > The input of the deep CNN model is RGB24 while the output is float for
> > each color channel. This is the filter's default behavior to output
> > format with gbrpf32le. And gbrp10le is also supported as the output,
> > so we can see the rendering result in a player, as a reference.
> >
> > To generate the model file, we need modify the original script a little.
> > - set name='y' for y_final within script at
> > https://github.com/gabrieleilertsen/hdrcnn/blob/master/network.py
> > - add the following code to the script at
> > https://github.com/gabrieleilertsen/hdrcnn/blob/master/hdrcnn_predict.
> > py
> >
> > graph = tf.graph_util.convert_variables_to_constants(sess,
> > sess.graph_def, ["y"]) tf.train.write_graph(graph, '.', 'graph.pb',
> > as_text=False)
> >
> > And I also uploaded the model file under
> https://drive.google.com/drive/folders/1URsRY5g-VdE-kHlP5vQoLoimMIZ-S
> X00?usp=sharing.
> >
> > The filter only works when tensorflow C api is supported in the
> > system, native backend is not supported since there are some different
> > types of layers in the deep CNN model, besides CONV and
> DEPTH_TO_SPACE.
> >
> > https://arxiv.org/pdf/1710.07480.pdf:
> >  author       = "Eilertsen, Gabriel and Kronander, Joel, and Denes,
> Gyorgy and Mantiuk, Rafał and Unger, Jonas",
> >  title        = "HDR image reconstruction from a single exposure using
> deep CNNs",
> >  journal      = "ACM Transactions on Graphics (TOG)",
> >  number       = "6",
> >  volume       = "36",
> >  articleno    = "178",
> >  year         = "2017"
> >
> > https://github.com/gabrieleilertsen/hdrcnn
> >
> > btw, as a whole solution, metadata should also be generated from the
> > sdr video, so to be encoded as a HDR video. Not supported yet.
> > This patch just focuses on this paper.
> >
> > This filter accepts 8bit frame (RGB24) and outputs 10bit/float frame,
> > and there's no reference image, so it is not feasible to use criteria such as
> PNSR, SSIM.
> >
> > I choose the same method described in the paper to demo the filter
> > effect, that means the frames before/after the filter are reduced by 3
> stops.
> >
> > The native video (test.native.mp4) is created from 7 png files at
> > https://github.com/gabrieleilertsen/hdrcnn/tree/master/data (the size
> > of the image is enlarged to 1920*1080 with extra area filled with white)
> with command line:
> > ffmpeg -f image2 -i ./img_%03d.png -c:v libx264 -preset veryslow -crf 1
> test.native.mp4.
> >
> > And two rgb24 videos are generated before/after the filter with -3
> > stops by modifying the code a little, see in the video folder at the
> > google drive (the same place as where the model file locates).
> >
> > For your convenient, I also dump png files from generated videos and
> > combine the before/after pngs into one file, see in png folder at the
> google drive.

I see three limitations from the code but haven't been noted in the texi or commit message:
1. Only support one resolution: 1920x1080. Other resolution can't been supported. 
2. RGB24 is the only input format can be supported. 
3. No meta data which may break encoder. 

(Should be good if can remove any of them).

> >
> > Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
> > ---
> > configure                |   1 +
> > doc/filters.texi         |  36 +++++++
> > libavfilter/Makefile     |   1 +
> > libavfilter/allfilters.c |   1 +
> > libavfilter/vf_sdr2hdr.c | 268
> > +++++++++++++++++++++++++++++++++++++++++++++++
> > 5 files changed, 307 insertions(+)
> > create mode 100644 libavfilter/vf_sdr2hdr.c
> >
> > diff --git a/configure b/configure
> > index b02b4cc..19138e8 100755
> > --- a/configure
> > +++ b/configure
> > @@ -3446,6 +3446,7 @@ sab_filter_deps="gpl swscale"
> > scale2ref_filter_deps="swscale"
> > scale_filter_deps="swscale"
> > scale_qsv_filter_deps="libmfx"
> > +sdr2hdr_filter_deps="libtensorflow"
> > select_filter_select="scene_sad"
> > sharpness_vaapi_filter_deps="vaapi"
> > showcqt_filter_deps="avcodec avformat swscale"
> > diff --git a/doc/filters.texi b/doc/filters.texi index
> > 0d9ff43..2e6a6af 100644
> > --- a/doc/filters.texi
> > +++ b/doc/filters.texi
> > @@ -14868,6 +14868,42 @@ Scale a subtitle stream (b) to match the main
> > video (a) in size before overlayin @end example @end itemize
> >
> > + at section sdr2hdr
> > +
> > +HDR image generation from a single exposure using deep CNNs with
> TensorFlow C library.
> > +
> > + at itemize
> > + at item
> > +paper:  see @url{https://arxiv.org/pdf/1710.07480.pdf}
> > +
> > + at item
> > +code with model and trained parameters: see
> > + at url{https://github.com/gabrieleilertsen/hdrcnn}
> > + at end itemize
> > +
> > +The filter accepts the following options:
> > +
> > + at table @option
> > +
> > + at item model_filename
> > +Set path to model file specifying network architecture and its
> > +parameters, can download from
> >
> + at url{https://drive.google.com/drive/folders/1URsRY5g-VdE-kHlP5vQoLoi
> m
> > +MIZ-SX00?usp=sharing}
> > +
> > + at item out_fmt
> > +the data format of the filter's output.
> > +
> > +It accepts the following values:
> > + at table @samp
> > + at item gbrpf32le
> > +force gbrpf32le output
> > +
> > + at item gbrp10le
> > +force gbrp10le output
> > + at end table
> > +
> > +Default value is @samp{gbrpf32le}.
> > +
> > + at end table
> > +
> > @anchor{selectivecolor}
> > @section selectivecolor
> >
> > diff --git a/libavfilter/Makefile b/libavfilter/Makefile index
> > 7c6fc83..936a525 100644
> > --- a/libavfilter/Makefile
> > +++ b/libavfilter/Makefile
> > @@ -365,6 +365,7 @@ OBJS-$(CONFIG_SOBEL_OPENCL_FILTER)
> += vf_convolution_opencl.o opencl.o
> > OBJS-$(CONFIG_SPLIT_FILTER)                  += split.o
> > OBJS-$(CONFIG_SPP_FILTER)                    += vf_spp.o
> > OBJS-$(CONFIG_SR_FILTER)                     += vf_sr.o
> > +OBJS-$(CONFIG_SDR2HDR_FILTER)                += vf_sdr2hdr.o
> > OBJS-$(CONFIG_SSIM_FILTER)                   += vf_ssim.o
> framesync.o
> > OBJS-$(CONFIG_STEREO3D_FILTER)               += vf_stereo3d.o
> > OBJS-$(CONFIG_STREAMSELECT_FILTER)           += f_streamselect.o
> framesync.o
> > diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c index
> > 484b080..622f9f3 100644
> > --- a/libavfilter/allfilters.c
> > +++ b/libavfilter/allfilters.c
> > @@ -322,6 +322,7 @@ extern AVFilter ff_vf_scale_npp; extern AVFilter
> > ff_vf_scale_qsv; extern AVFilter ff_vf_scale_vaapi; extern AVFilter
> > ff_vf_scale2ref;
> > +extern AVFilter ff_vf_sdr2hdr;
> > extern AVFilter ff_vf_select;
> > extern AVFilter ff_vf_selectivecolor;
> > extern AVFilter ff_vf_sendcmd;
> > diff --git a/libavfilter/vf_sdr2hdr.c b/libavfilter/vf_sdr2hdr.c new
> > file mode 100644 index 0000000..85a58ea
> > --- /dev/null
> > +++ b/libavfilter/vf_sdr2hdr.c
> > @@ -0,0 +1,268 @@
> > +/*
> > + * Copyright (c) 2018 Guo Yejun
> > + *
> > + * 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
> > + * Filter implementing HDR image generation from a single exposure using
> deep CNNs.
> > + * https://arxiv.org/pdf/1710.07480.pdf
> > + */
> > +
> > +#include "avfilter.h"
> > +#include "formats.h"
> > +#include "internal.h"
> > +#include "libavutil/opt.h"
> > +#include "libavutil/qsort.h"
> > +#include "libavformat/avio.h"
> > +#include "libswscale/swscale.h"
> > +#include "dnn_interface.h"
> > +#include <math.h>
> > +
> > +typedef struct SDR2HDRContext {
> > +    const AVClass *class;
> > +
> > +    char* model_filename;
> > +    enum AVPixelFormat out_fmt;
> > +    DNNModule* dnn_module;
> > +    DNNModel* model;
> > +    DNNData input, output;
> > +} SDR2HDRContext;
> > +
> > +#define OFFSET(x) offsetof(SDR2HDRContext, x) #define FLAGS
> > +AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM static
> const
> > +AVOption sdr2hdr_options[] = {
> > +    { "model_filename", "path to model file specifying network
> architecture and its parameters", OFFSET(model_filename),
> AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
> > +    { "out_fmt", "the data format of the filter's output, it could be
> gbrpf32le [default] or gbrp10le", OFFSET(out_fmt),
> AV_OPT_TYPE_PIXEL_FMT, {.i64=AV_PIX_FMT_GBRPF32LE},
> AV_PIX_FMT_NONE, AV_PIX_FMT_NB - 1, FLAGS },
> > +    { NULL }
> > +};
> > +
> > +AVFILTER_DEFINE_CLASS(sdr2hdr);
> > +
> > +static av_cold int init(AVFilterContext* context) {
> > +    SDR2HDRContext* ctx = context->priv;
> > +
> > +    if (ctx->out_fmt != AV_PIX_FMT_GBRPF32LE && ctx->out_fmt !=
> AV_PIX_FMT_GBRP10LE) {
> > +        av_log(context, AV_LOG_ERROR, "could not support the output
> format\n");
> > +        return AVERROR(ENOSYS);
> > +    }
> > +
> > +    ctx->dnn_module = ff_get_dnn_module(DNN_TF);
> > +    if (!ctx->dnn_module){
> > +        av_log(context, AV_LOG_ERROR, "could not create DNN
> module for tensorflow backend\n");
> > +        return AVERROR(ENOMEM);
> > +    }
> > +    if (!ctx->model_filename){
> > +        av_log(context, AV_LOG_ERROR, "model file for network was
> not specified\n");
> > +        return AVERROR(EIO);
> > +    }
> > +    if (!ctx->dnn_module->load_model) {
> > +        av_log(context, AV_LOG_ERROR, "load_model for network was
> not specified\n");
> > +        return AVERROR(EIO);
> > +    }
> > +    ctx->model =
> (ctx->dnn_module->load_model)(ctx->model_filename);
> > +    if (!ctx->model){
> > +        av_log(context, AV_LOG_ERROR, "could not load DNN
> model\n");
> > +        return AVERROR(EIO);
> > +    }
> > +    return 0;
> > +}
> > +
> > +static int query_formats(AVFilterContext* context) {
> > +    const enum AVPixelFormat in_formats[] = {AV_PIX_FMT_RGB24,
> > +
> AV_PIX_FMT_NONE};
> > +    enum AVPixelFormat out_formats[2];
> > +    SDR2HDRContext* ctx = context->priv;
> > +    AVFilterFormats* formats_list;
> > +    int ret = 0;
> > +
> > +    formats_list = ff_make_format_list(in_formats);
> > +    if ((ret = ff_formats_ref(formats_list,
> &context->inputs[0]->out_formats)) < 0)
> > +        return ret;
> > +
> > +    out_formats[0] = ctx->out_fmt;
> > +    out_formats[1] = AV_PIX_FMT_NONE;
> > +    formats_list = ff_make_format_list(out_formats);
> > +    if ((ret = ff_formats_ref(formats_list,
> &context->outputs[0]->in_formats)) < 0)
> > +        return ret;
> > +
> > +    return 0;
> > +}
> > +
> > +static int config_props(AVFilterLink* inlink) {
> > +    AVFilterContext* context = inlink->dst;
> > +    SDR2HDRContext* ctx = context->priv;
> > +    AVFilterLink* outlink = context->outputs[0];
> > +    DNNReturnType result;
> > +
> > +    // the dnn model is tied with resolution due to deconv layer of
> tensorflow
> > +    // now just support 1920*1080 and so the magic numbers within this
> file
> > +    if (inlink->w != 1920 || inlink->h != 1080) {
> > +        av_log(context, AV_LOG_ERROR, "only support frame size with
> 1920*1080\n");
> > +        return AVERROR(ENOSYS);
> > +     }
> > +
> > +    ctx->input.width = 1920;
> > +    ctx->input.height = 1088;  //the model requires height is a multiple
> of 32,

Would be better to avoid any hard code. I prefer something like:

ctx->input.width = inlink->w;
ctx->input.height = FFALIGN(inlink->h, 32);


> > +    ctx->input.channels = 3;
> > +
> > +    result = (ctx->model->set_input_output)(ctx->model->model,
> &ctx->input, &ctx->output);
> > +    if (result != DNN_SUCCESS){
> > +        av_log(context, AV_LOG_ERROR, "could not set input and
> output for the model\n");
> > +        return AVERROR(EIO);
> > +    }
> > +
> > +    memset(ctx->input.data, 0, ctx->input.channels * ctx->input.width *
> ctx->input.height * sizeof(float));
> > +    outlink->h = 1080;
> > +    outlink->w = 1920;

And also here:
outlink->h = inlink->h;
outlink->w = inlink->w;

Then one more resolution supported, there is few code need to be changed.

> > +    return 0;
> > +}
> > +
> > +static float qsort_comparison_function_float(const void *a, const
> > +void *b) {
> > +    return *(const float *)a - *(const float *)b; }
> > +
> > +static int filter_frame(AVFilterLink* inlink, AVFrame* in) {
> > +    DNNReturnType dnn_result = DNN_SUCCESS;
> > +    AVFilterContext* context = inlink->dst;
> > +    SDR2HDRContext* ctx = context->priv;
> > +    AVFilterLink* outlink = context->outputs[0];
> > +    AVFrame* out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
> > +    int total_pixels = in->height * in->width;
> > +
> > +    if (!out){
> > +        av_log(context, AV_LOG_ERROR, "could not allocate memory
> for output frame\n");
> > +        av_frame_free(&in);
> > +        return AVERROR(ENOMEM);
> > +    }
> > +
> > +    av_frame_copy_props(out, in);
> > +
> > +    for (int i = 0; i < in->linesize[0] * in->height; ++i) {
> > +        ctx->input.data[i] = in->data[0][i] / 255.0f;
> > +    }
> > +
> > +    dnn_result = (ctx->dnn_module->execute_model)(ctx->model);
> > +    if (dnn_result != DNN_SUCCESS){
> > +        av_log(context, AV_LOG_ERROR, "failed to execute loaded
> model\n");
> > +        return AVERROR(EIO);
> > +    }
> > +
> > +    if (ctx->out_fmt == AV_PIX_FMT_GBRPF32LE) {
> > +        float* outg = (float*)out->data[0];
> > +        float* outb = (float*)out->data[1];
> > +        float* outr = (float*)out->data[2];
> > +        for (int i = 0; i < total_pixels; ++i) {
> > +            float r = ctx->output.data[i*3];
> > +            float g = ctx->output.data[i*3+1];
> > +            float b = ctx->output.data[i*3+2];
> > +            outr[i] = r;
> > +            outg[i] = g;
> > +            outb[i] = b;
> > +        }
> > +    } else 

Would be better to change to "else if (fmt=gbrp10le)", and give an assert in the below? 
(I believe the format should be checked again though it has been checked in the initialization stage)

{
> > +        // here, we just use a rough mapping to the 10bit contents
> > +        // meta data generation for HDR video encoding is not
> supported yet
> > +        float* converted_data = (float*)av_malloc(total_pixels * 3 *
> sizeof(float));
> > +        int16_t* outg = (int16_t*)out->data[0];
> > +        int16_t* outb = (int16_t*)out->data[1];
> > +        int16_t* outr = (int16_t*)out->data[2];
> > +
> > +        float max = 1.0f;
> > +        for (int i = 0; i < total_pixels * 3; ++i) {
> > +            float d = ctx->output.data[i];
> > +            d = sqrt(d);
> > +            converted_data[i] = d;
> > +            max = FFMAX(d, max);
> > +        }
> > +
> > +        if (max > 1.0f) {
> > +            AV_QSORT(converted_data, total_pixels * 3, float,
> qsort_comparison_function_float);
> > +            // 0.5% pixels are clipped
> > +            max = converted_data[(int)(total_pixels * 3 * 0.995)];
> > +            max = FFMAX(max, 1.0f);
> > +
> > +            for (int i = 0; i < total_pixels * 3; ++i) {
> > +                float d = ctx->output.data[i];
> > +                d = sqrt(d);
> > +                d = FFMIN(d, max);
> > +                converted_data[i] = d;
> > +            }
> > +        }
> > +
> > +        for (int i = 0; i < total_pixels; ++i) {
> > +            float r = converted_data[i*3];
> > +            float g = converted_data[i*3+1];
> > +            float b = converted_data[i*3+2];
> > +            outr[i] = r / max * 1023;
> > +            outg[i] = g / max * 1023;
> > +            outb[i] = b / max * 1023;
> > +        }
> > +
> > +        av_free(converted_data);
> > +    }
> > +
> > +    av_frame_free(&in);
> > +    return ff_filter_frame(outlink, out); }
> > +
> > +static av_cold void uninit(AVFilterContext* context) {
> > +    SDR2HDRContext* ctx = context->priv;
> > +
> > +    if (ctx->dnn_module){
> > +        (ctx->dnn_module->free_model)(&ctx->model);
> > +        av_freep(&ctx->dnn_module);
> > +    }
> > +}
> > +
> > +static const AVFilterPad sdr2hdr_inputs[] = {
> > +    {
> > +        .name         = "default",
> > +        .type         = AVMEDIA_TYPE_VIDEO,
> > +        .config_props = config_props,
> > +        .filter_frame = filter_frame,
> > +    },
> > +    { NULL }
> > +};
> > +
> > +static const AVFilterPad sdr2hdr_outputs[] = {
> > +    {
> > +        .name = "default",
> > +        .type = AVMEDIA_TYPE_VIDEO,
> > +    },
> > +    { NULL }
> > +};
> > +
> > +AVFilter ff_vf_sdr2hdr = {
> > +    .name          = "sdr2hdr",
> > +    .description   = NULL_IF_CONFIG_SMALL("HDR image generation
> from a single exposure using deep CNNs."),
> > +    .priv_size     = sizeof(SDR2HDRContext),
> > +    .init          = init,
> > +    .uninit        = uninit,
> > +    .query_formats = query_formats,
> > +    .inputs        = sdr2hdr_inputs,
> > +    .outputs       = sdr2hdr_outputs,
> > +    .priv_class    = &sdr2hdr_class,
> > +    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,
> > +};
> > --
> > 2.7.4
> >
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> I have tested this patch, have some question:
> 
> 1. It must use Tensorflow, why don’t support native mode?
> 2. It is must input 1920x1080 resolution, can this add more resolution
> support?

Yup, would be better if we can support more resolution. 

> 3. i looked into the project of hdrcnn, that License is BSD-3-Clause, Is this will
> have License problem?

Should be not a problem since it hasn't involved any source code of this project.
> 
> Thanks
> 
> Steven
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel


More information about the ffmpeg-devel mailing list