I really like how the paper proposes a different approach for wide-baseline matching; viewing it as a classification problem.
I'm not that familiar with classification methods, but the author mentions PCA and Randomized Trees along with kernel PCA as classification techniques, however, he doesn't provide any performance comparisons between Randomized Trees and PCA or any other classification technique such as k-nearest neighbor, svm, or neural networks.
Is Randomized Tree's advantage over these other techniques obvious?
And, are there any other classification techniques currently being used for wide-baseline matching?

By request, here is a conference paper by Lepetit et al. that lead to the journal article that's linked on the 252c site: http://cvlab.epfl.ch/publications/publicat...005/LepetitLF05.pdf

I think it is very interesting that the training set is potentially infinite since they synthesize "view sets".

The authors mention that they use their own "simple and fast" keypoint detector and that it suffices for operation under even large perspective and scale variations. It would be worthwhile to plug in other simple keypoint detectors to see if a potentially more robust one exists. Also, while it would be more computationally intensive, it would be interesting to see if and how much using a more sophisticated keypoint detector improves the detection and tracking performance of their system.

I thought the way they synthesized a model to generate a large training set was interesting. More specifically, the way they added noise to an image patch and composited the image patch over a noisy background to improve realism and thus improve keypoint classification under large positional shifts with cluttered backgrounds (page 9). Do other methods do this (adding white noise to improve performance), or is this unique to the authors' method?
Also the authors compared their method to SIFT. It would be interesting to see a comparison to SURF or other improved methods.