Computational demanding sections about openbr

[jklontz]

Nope! Sounds great. If you haven't already, I would suggest benchmarking the example application face_recognition_evaluation to get a sense as to what's computationally demanding. The two routines I'd like to focus on accelerating are br::Enroll and br::Compare. I think OpenMP will be a dead end in this case as we already distribute the workload across the CPU quite efficiently using thread pools. MPI support should be relatively straightforward to implement. GPU work is a bit trickier but perhaps more interesting and more beneficial in the long run. The good news is we mostly make calls into OpenCV for our computations, and many of the calls we make have GPU equivalents. The tricky part will be finding the right software abstraction to allow algorithms to default back to the CPU if a GPU isn't available, and also avoid memory transfers to/from the device when multiple GPU operations should happen in series.

On a somewhat related note, this seems like a good opportunity to pitch my side project www.openbiometrics.org/likely which I'm hoping will be the long term replacement for OpenCV and obviate many of these considerations :)

Anyway, please follow up with what specific avenue you'd like to pursue and I can share more thoughts about how it should be incorporated. Thanks for your interest in the project!

[jklontz]

Closed due to inactivity.