Opinion about implementation feasibility (Patch Match algorithm)

[Filippos_Kokkinos]

Hello guys!

I am thinking to create the Patch Match algorithm using Pytorch. The idea behind the implementation is to use Tensors with GPU support in order to do the calculations fast. A fast overview of patch match algorithm is the following:
Patch match is a randomized algorithm that tries to approximate nearest neighbors of a reference patch in a target set of patches. Here is a video explanation: https://www.youtube.com/watch?v=fMe19oTz6vk (4 min)

Do you think that a Pytorch implementation will be feasible? Also, this algorithm make heavy use of indices so I am not sure if the end result will be fast.

Best regards,
Filippos

[tom (Thomas V)]

Hello Filippos,

it would sound like it is feasible (note that there also is a PatchMatch wikipedia page). A month ago, Francisco Massa suggested a possible approach for implementation that should be relatively easy to follow:

Need urgent help to implement Patch Match in Pytorch

As I mentioned on Twitter, I’d start with something based on numpy and adapt/optimize for pytorch. Something like https://github.com/jabooth/patchmatch/blob/master/patchmatch.py for example. It will be definitely slower than using hand-made kernels, but will definitely be easier to debug. Good luck!

Of course, we’d love to see your implementation once it is done.

Best regards

Thomas