[FFmpeg-devel] [PATCH] libavfilter: Removes stored DNN models. Adds support for native backend model file format in tf backend. Removes scaling and conversion with libswscale and replaces input formats with AV_PIX_FMT_GRAYF32 in sr filter

Pedro Arthur bygrandao at gmail.com
Mon Sep 10 20:24:07 EEST 2018

2018-09-06 8:44 GMT-03:00 Sergey Lavrushkin <dualfal at gmail.com>:

> Here is the patch with reverted changes on sws removal. I didn't split the
> patch into two patches, because code, that supports native model file
> format in
> tf, is partially from code of default model construction, which is removed
> with default
> models and stored data.

The scale_factor option is not necessary, it should be stored in the model
file. As the weights are trained for a specific factor, using anything
different from that will give bad results.
It seems the native depth to space conversion is buggy a few lines in the
top of output image are duplicated, tf backend is ok. BTW this bug was not
introduced by this patch.

Other than that LGTM, the above fixes can be done in separated patches.
I may push it by Friday.

More information about the ffmpeg-devel mailing list