I'm working with VLAD codes, and their precision highly depends by the descriptor used.
In order of precision, I obtained the best results with:
- This hessian-affine + sift descriptor implementation
- SURF by OpenCV
- SIFT by OpenCV
Surprisingly, SIFT obtained worse performance (both in time and precision) than SURF. However, there is no match with hessaff: it results are much much better than the others to (map=0.59 against 0.49 of surf). However, it's very slow too. Compute hesaff takes almost 10 times more than SURF.
For this reason, I would like to parallelize it and make it fast as SURF. Finally, I want to propose it as an OpenCV contribution, I think that a lot of people would appreciate it. Notice that I have not much knowledge in Computer Vision and Image Processing, but a lot in parallel systems.
However, I never contributed to OpenCV and I wonder if this is a good idea or not. For example, since this code is pretty much old, they implement SIFT detection and description from scratch. I don't know how well this could be accepted by OpenCV (since SIFT is already implemented). What do you suggest me to do in this scenario?