FAST/ORB version comparison with OpenCV for acceleration (opencl/cuda)

[SteffenEhrle]

Hi there,

I am experimenting with the opencv versions of FAST/ORB to speed up feature extraction time.
I'd like to add (if possible) opencl and cuda versions of the whole preprocessing step to stella_vslam.

looking into orb_extractor.cc, I find FAST being executed parallel on overlapping patches.

What is the difference to opencv::ORB implementation (which also uses a image pyramid)?

[SteffenEhrle]

CPU opencv ORB:

OpenCL opencv ORB:

Cuda opencv ORB:

stella_VSLAM ORB:

Ok there is a clear difference concerning number and distribution of features.