avfilter: New scene classification filter

configure

@@ -3581,6 +3581,7 @@ interlace_filter_deps="gpl"
 kerndeint_filter_deps="gpl"
 ladspa_filter_deps="ladspa libdl"
 lensfun_filter_deps="liblensfun version3"
+livepeer_dnn_filter_select="dnn"
 lv2_filter_deps="lv2"
 mcdeint_filter_deps="avcodec gpl"
 movie_filter_deps="avcodec avformat"

doc/filters.texi

@@ -14362,6 +14362,73 @@ lut2='if(lt(x,y),0,if(gt(x,y),pow(2,bdx)-1,pow(2,bdx-1))):if(lt(x,y),0,if(gt(x,y
 @end example
 @end itemize
 
+@section livepeer_dnn
+
+
+Apply convolutional neural network based methods on the input video to classify
+the overall scene, for example classifying input frames as Soccer or Adult content.
+Supported models: TensorFlow model files(.pb)
+
+The filter accepts the following options:
+
+@table @option
+@item dnn_backend
+Specify which DNN backend to use for model loading and execution. This option accepts
+the following values:
+
+@table @samp
+@item native
+Native implementation of DNN loading and execution. Native implementation is under testing, so tensorflow backend is recommended.
+
+@item tensorflow
+TensorFlow backend. To enable this backend you
+need to install the TensorFlow for C library (see
+@url{https://www.tensorflow.org/install/install_c}) and configure FFmpeg with
+@code{--enable-libtensorflow}
+@end table
+
+Default value is @samp{tensorflow}.
+
+@item backend_configs
+Set the configuration passed to the DNN backend. Useful to specify GPU ID for tensorflow backend.
+
+@item model
+Set path to model file specifying network architecture and its parameters.
+Note that different backends use different file formats. TensorFlow can load files for only its format.
+
+@item input
+Set the input name of the dnn network.
+
+@item output
+Set the output name of the dnn network.
+
+@item logfile
+Set the output log file path for save inference results.
+
+@end table
+
+@itemize
+@item
+Software transcoding for every 10th frame:
+@example
+ffmpeg -i input.mp4 -vf select='not(mod(n\,10))',livepeer_dnn=model=tasmodel.pb:input=input_1:output=reshape_3/Reshape:logfile=outlog.txt out.mp4
+@end example
+
+@item
+Hardware transcoding for every 10th frame:
+@example
+ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 -vf select='not(mod(n\,10))',hwdownload,format=pix_fmts=nv12,livepeer_dnn=model=tasmodel.pb:input=input_1:output=reshape_3/Reshape:logfile=outlog.txt -c:v h264_nvenc out.mp4
+@end example
+
+@item
+ffprobe example:
+@example
+ffprobe -show_entries frame_tags=lavfi.lvpdnn.text -f lavfi -i "movie=input.mp4,livepeer_dnn=model=tasmodel.pb:input=input_1:output=reshape_3/Reshape"
+@end example
+
+@end itemize
+
+
 @section maskedclamp
 
 Clamp the first input stream with the second input and third input stream.

libavfilter/Makefile

@@ -312,6 +312,7 @@ OBJS-$(CONFIG_LENSCORRECTION_FILTER)         += vf_lenscorrection.o
 OBJS-$(CONFIG_LENSFUN_FILTER)                += vf_lensfun.o
 OBJS-$(CONFIG_LIBVMAF_FILTER)                += vf_libvmaf.o framesync.o
 OBJS-$(CONFIG_LIMITER_FILTER)                += vf_limiter.o
+OBJS-$(CONFIG_LIVEPEER_DNN_FILTER)           += vf_livepeer_dnn.o
 OBJS-$(CONFIG_LOOP_FILTER)                   += f_loop.o
 OBJS-$(CONFIG_LUMAKEY_FILTER)                += vf_lumakey.o
 OBJS-$(CONFIG_LUT1D_FILTER)                  += vf_lut3d.o

libavfilter/allfilters.c

@@ -296,6 +296,7 @@ extern AVFilter ff_vf_lenscorrection;
 extern AVFilter ff_vf_lensfun;
 extern AVFilter ff_vf_libvmaf;
 extern AVFilter ff_vf_limiter;
+extern AVFilter ff_vf_livepeer_dnn;
 extern AVFilter ff_vf_loop;
 extern AVFilter ff_vf_lumakey;
 extern AVFilter ff_vf_lut;

libavfilter/dnn/dnn_backend_tf.c

@@ -153,7 +153,7 @@ static DNNReturnType get_input_tf(void *model, DNNData *input, const char *input
     TF_DeleteStatus(status);
 
     // currently only NHWC is supported
-    av_assert0(dims[0] == 1);
+    av_assert0(dims[0] == 1 || dims[0] == -1);
     input->height = dims[1];
     input->width = dims[2];
     input->channels = dims[3];