[IPOL discuss] optic flow algorithms in ipol

[Daniel Kondermann]

> Otherwise, an unpublished workshop comparing many implementations is
> interesting as a tool.  Indeed, all the implementations of this
> algorithm that I tried, including my own, produce vastly different
> results.

I made the same experience. Deqing Sun also wrote a paper addressing
some of these issues:
http://www.cs.brown.edu/~dqsun/pubs/cvpr_2010_flow.pdf

When strictly talking about H&S, there is an implementation given in the
paper itself. It would be interesting to have such a *really* strict
implementation.
For example, our implementation uses a different discretization of the
laplace-operator and a completely different optimization scheme. We also
experimented with various ways to compute image derivatives as well as
pyramids. So there are a lot of H&S-like implementations but I have not
yet heard of an exact implementation.

Generally speaking, the problem is that modern publications almost never
contain all implementation details.
In my opinion deciding whether another implementation of the same
"paper" (rather than "algorithm") should be published, could therefore
be based on the difference in the raw results when the same input data
and parameters are used.

Best,
Daniel

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 552 bytes
Desc: OpenPGP digital signature
URL: <http://tools.ipol.im/mailman/archive/discuss/attachments/20110927/bbc44866/attachment.pgp>