[daala] Fwd: Great encoder idea: Face recognition & Double bitrate

[Alexey Eromenko]

Hi,

I did some encoding of Skyrim video clip at very low bitrates (200
kbps, where both H.264 and VP8 can't keep up) and the obvious outcome,
is that faces deserve higher bitrate, because it's the first thing
humans look at.

For example, my Galaxy S3 smartphone can detect faces in most photos.
(so it can be done)

This idea is simple: The encoder must analyze each frame with a "face
detector", and if a macroblock contains a face, then give it higher
bitrate (double? quadruple? bitrate), while other objects (such as the
carpet or a wall) can be given lower (half?) the bitrate.
This is all about human perception, after all.

This encoder trick should help improve quality with all bitrates, but
will kick ass and become especially obvious at low bitrates.
The downside, is that encoder will become slightly slower.

Examples: (Skyrim: Face + background)
200 kbps
https://docs.google.com/file/d/0BycgkMZbeQOzaHE2T2tLZjJCZ2c/edit?usp=sharing
400 kbps
https://docs.google.com/file/d/0BycgkMZbeQOzd2ZlbGptMzkzS1k/edit?usp=sharing
600 kbps
https://docs.google.com/file/d/0BycgkMZbeQOzM3Y2c3RlQ2FiYkk/edit?usp=sharing

What do you think of it ?
--
-Alexey Eromenko "Technologov"