Is Approximate Nearest Neighbour the fastest feature matching in Computer Vision?

Question

When using feature descriptors [like SIFT, SURF] - is Approximate Nearest Neighbour the fastest method to do matching between images?

Answer 1

I'd say that Euclidean distnace based nearest neighbor would be the easiest to implement, but not necessarily the fastest.

I'd agree that approximate nearest neighbor or 'best bin first' would be the quickest at identifying which image in your background set most closely resembles the probe image.

If your trying to identify a single object in the image, things will be a little more difficult.

Answer 2

You should check out pyramid match kernel, which is one of the most successful algorithms for image matching with local features so far. It has a linear time complexity, as opposed to comparing every feature in image A to every feature in image B, which is O(n^2). There is also a free implementation.