Please post your thoughts about SIFT here.

SIFT seems like a pretty clever technique, but in reading the paper, I felt like they just pulled a bunch of values out of a hat. They claim invariance to all sorts of phenomena, like rotation and illumination, but they don't really present proofs. I hear that SIFT is popular among some groups, so it's probably worth something. Still, I'd like to have a better understanding of the underlying reasoning for such features, not just values of sigma.

After working with the algorithm for a while now I'm pretty impressed with its performance. Nevertheless there's a sense in which its a bit TOO discriminative for object recognition. Lowe's illumination example aside, a small change in lighting can easily thwart recognition. Is SIFT really summarizing essential characteristics of an object, or simply identifying a few "accidental" points that happen to be descriptive?

troceltdar