[MEncoder-users] motion estimation across different codecs

[Reimar Döffinger]

On Mon, Mar 05, 2012 at 06:14:50PM +0100, Adrian Ulges wrote:
> I observed that the distribution of the motion compensation vectors
> across both sources is very different (which may not be surprising):
> Videos from source 2 tend to have a lot more purely
> horizontal/vertical motion vectors (i.e., either v_x or v_y is usually
> zero), while videos from source 2 contain more diverse motion vectors.
> This is bad, because I want to do some fancy machine learning on these
> motion vectors.

You haven't said why you would think they should be similar.
Anyway, for video encoding the point of the motion vectors is to
minimize encoding cost.
There are slightly different algorithm in use (e.g. in frame
interpolation filters) when the goal is to detect motion.
Still there are a lot of motion search algorithms available and
you can try different ones.
Though xvid is an external project and this is not really
the best place to ask about it, and I suspect that the FFmpeg
encoder actually gives you a wider choice of algorithms.