flannIndex in openCV is designed for matching 2 images through binary descriptors.
Anyway LSH is heavily used in CBIR in order not to "comparing two images" but "find the most similar image in the dataset", which is obviously something different.
How we can implement such a thing in OpenCV?
