[FFmpeg-user] minterpolate ...for Paul

pdr0 pdr0 at shaw.ca
Fri Jan 29 03:01:14 EET 2021

Mark Filipak (ffmpeg) wrote
> I've never heard of "optical flow errors". What could they be? (Got any
> links to 
> explanations?)

The artifacts in your video are optical flow errors :)

If you've ever used it - you'd recognize these artifacts. There are very

There are about a dozen prototypical "fail" categories or common errors that
plague all types of optical flow

These are errors either of motion vectors, or object flow (object boundaries
or "masks"), occlusion errors. 

Internet is full of examples, explanations. The topic is rather large, just
search google, there is lots of info. If you have a specific question then

Sometimes you get clean interpolated frame results;  but sometimes there are
massive distracting errors. It varies by situation and sources. 

Your example has one of the common categories of "fail" where there are
repeating patterns and textures. It falls under the "Picket Fence" fail . A
prototypical tracking or dolly shot by a picket fence, or brick wall will
come up with interpolation errors

The peripheral edges error are common because there is less data beyond the
periphery of the frame, for n-1, n+1 and the motion vectors are less
accurate compared to the center of the frame

Another common one is when objects pass over another. The flow masks aren't
perfect and you end up with blobby edge artifacts around objects

>>...For artifacts around frame edges, letterbox edges usually some form
>> of padding is used. I don't think ffmpeg minterpolate has those.
> I've done that. The result was just okay. The slight riffling on the frame
> boundaries during camera 
> panning isn't all that objectionable to me. It occurs to me that
> minterpolute could queue frames and 
> look 'forward' to later frames in order to resolve boundary macroblock
> artifacts -- afterall, it has 
> the motion vectors, eh?

Some algorithms can use N-3, N-2, N-1, N, N+1, N+2, N+3, I don't think
minterpolate can.  More is not always better. Often you get more
contamination with a larger "window"

Sometimes just changing the blocksize can produce better (or worse) results.
The problem is ffmpeg minterpolate is soooo slow, and you have no usable
preview. Some of the other methods mentioned earlier do have previews -  so
you can tweak settings, preview, readjust etc....

Sent from: http://ffmpeg-users.933282.n4.nabble.com/

More information about the ffmpeg-user mailing list