[FFmpeg-devel] [PATCH V3] Add a filter implementing HDR image reconstruction from a single exposure using deep CNNs
Guo, Yejun
yejun.guo at intel.com
Mon Oct 22 07:19:53 EEST 2018
From: Vittorio Giovara [mailto:vittorio.giovara at gmail.com]
Sent: Friday, October 19, 2018 11:33 PM
To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
Cc: Guo, Yejun <yejun.guo at intel.com>; Guo at ffbox0-bg.ffmpeg.org
Subject: Re: [FFmpeg-devel] [PATCH V3] Add a filter implementing HDR image reconstruction from a single exposure using deep CNNs
On Fri, Oct 19, 2018 at 10:11 AM Guo, Yejun <yejun.guo at intel.com<mailto:yejun.guo at intel.com>> wrote:
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
> can you add some usage documentation to doc/filters.texi?
sure, will add it.
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)
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.
> Is this something you are working on and will it be added later?
yes, this is in our team’s todo list.
v3: use int16_t instead of short
v2: use AV_OPT_TYPE_PIXEL_FMT for filter option
remove some unnecessary code
Use in->linesize[0] and FFMAX/FFMIN
remove flag AVFILTER_FLAG_SLICE_THREADS
add av_log message when error
> there is no need for this block to be left in the commit log
ok, will remove it.
Signed-off-by: Guo, Yejun <yejun.guo at intel.com<mailto:yejun.guo at intel.com>>
---
libavfilter/Makefile | 1 +
libavfilter/allfilters.c | 1 +
libavfilter/vf_sdr2hdr.c | 266 +++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 268 insertions(+)
create mode 100644 libavfilter/vf_sdr2hdr.c
+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);
+ }
+
+#if (CONFIG_LIBTENSORFLOW == 1)
+ 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;
+#else
+ return AVERROR(EIO);
+#endif
+}
> this is incorrect, what you should do is make libtensorflow a dependency of this filter in the configure file and disable this filter when it is not enabled
thanks, will fix it.
+
+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);
+ }
> is there any work planned to extend this to other resolutions?
yes, the plan is to fix tensorflow first, to make deconv layer not tied with resolution. it is also in our team’s todo list.
+
+ ctx->input.width = 1920;
+ ctx->input.height = 1088; //the model requires height is a multiple of 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;
+ 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;
+
+ av_frame_copy_props(out, in);
> check for allocation failures here
thanks, will add the check.
+
+ 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 {
+ // 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*)malloc(total_pixels * 3 * sizeof(float));
> don't use malloc, replace with av_malloc, same for free below
thanks, will fix it.
+ 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;
+ }
+
+ 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 reconstruction from a single exposure using deep CNNs."),
> why "reconstruction"? there is nothing to construct back if the source wasn't hdr to begin with
> "tonemap" is probably a better term here, in my opinion
> same for previous uses
there is more detail data generated with the dnn model, the model accepts sdr frame and generates hdr data.
see more detail in paper @https://arxiv.org/pdf/1710.07480.pdf, and the description comes from the title of this paper.
--
Vittorio
More information about the ffmpeg-devel
mailing list