Binary descriptors about vocabtree2

[snavely]

Hi Oleg,

You are right that binary descriptors are a whole different kind of beast.
Is there a particular reason you want to go with them? Speed / memory?
Anyhow, maybe start with looking at Vincent Lepetit's work -- he's looked
at creating binary descriptors as well as matching them. (
http://cvlabwww.epfl.ch/~lepetit/). Also take a look at this paper:
http://www.cs.ubc.ca/~lowe/papers/12mujaCRV.pdf.

Best,
Noah

On Thu, Dec 4, 2014 at 9:10 PM, Oleg Alexandrov [email protected]
wrote:

Noah,

VocabTree2 works very well for me with SIFT descriptors, as it was
designed. I am looking for a solution using binary descriptors, such as
BRISK. From what I've been reading, it appears that VocabTree2 may not
generalize well to binary descriptors since it uses k-means and Euclidean
distance, while for binary descriptors the Hamming distance and other ways
of clustering may be more appropriate. I wonder if you have any insights on
this, and if by chance any pointers to relevant software. Thank you.

—
Reply to this email directly or view it on GitHub
#5.

[oleg-alexandrov]

Noah, thank you, this is very helpful! In both places people compare algorithms against locally sensitive hashing, which is in FLANN and OpenCV. I'll try to dig along those lines.

Indeed, we care a lot about speed and memory. Our application will be a SLAM-like algorithm on a mobile device where things are pretty constrained.