Hi.
There are currently 3 adaptive quatizer options available from the command line.
AQ 3 is an experimental option for live video conferencing so probablyy not what you want.
AQ1 adpats the quantizer within frames based upon the spatial variance of each region. Low variance regions get a lower quantizer than high variance regions.
AQ2 adapts the quantizer using as its main trigger the actual number of bits spent coding each region or block, relative to a target average for the frame.
Both AQ 1 and 2 at the moment only update the segmentation map once for each group of frames (defined by the positioning of alternate reference frames) rather than every frame. This is to reduce the overhead and works on the assumption that the map will usuaully remain quite stable for a small group of frames.
AQ1 and AQ 2 both give a strong improvment in psychovisual quality for some content (for example "ParkJoy" from the derf set). They almost always hurt metrics such as PSNR, but in some cases help the SSIM metric.
We are continuing to evaluate AQ methods and their impact on pschovisual and objective metrics over a wider corpus of clips.
I hope this helps.
Paul Wilkins