Keypoint Recognition using Randomized Trees, V. Lepetit and P. Fua, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 28, Nr. 9, pp. 1465--1479, 2006.
Using randomized trees for feature indexing. A more extreme view of the "simple features" ideas.
Subscribe to:
Post Comments (Atom)
Hey,
ReplyDeletethe link above actually points to the paper
'Fast Keypoint Recognition using Random Ferns' which will be the main focus of the talk Wednesday.
If you are interested you can find the paper 'Keypoint Recognition using Randomized Trees' at the following link: http://cvlab.epfl.ch/publications/publications/2006/LepetitF06.pdf
My concern about these two approaches (both randomized trees and ferns) is what kind of localization accuracy they offer for 3D recognition. Localization accuracy is an often overlooked aspect of invariant descriptors, while e.g. being the main reason why SURF is a BAD idea when doing 3D recognition/pose recovery (and an important reason why SIFT is used when accurate placement of keypoints is required). A recent study presented in ICRA 09 (http://www2.cvl.isy.liu.se/~perfo/papers/viksten_icra09.pdf) compared a multitude of different descriptors for 6D pose estimation, unsurprisingly concluding that their own descriptor was best, SIFT was second and SURF performed pretty terribly. Randomized trees and Ferns were sadly not included in the comparison, but it would be interesting to see how these two would perform.
ReplyDelete