I'm trying to build an CBIR (content based image retrieval) project and for the first step I'm looking at a good feature. This should be invariant on several cues: scale rotation illumination and device, and I've not any database or previous knowledge of the scene. Up to now I've read several papers and it seems a very difficult goal. In general, the common feature is based on colors, but these algorithms has the prior that all scenes are taken with same device models, or has an heavy training step.. or other simplifications. In my case I have different devices so that these devices uses different color representation (due to different acquisition, quantization etc..). After color histograms i've tried with SIFT but the matching of these fails even in the easiest cases (also this due to different devices..) Just read the abstract of this paper daunted me a bit :)
It is clear that these images need some "normalization" preprocessing... first of all i put all images in the same dimensions (according to the littlest)... then i need something that make colours appear similar.. according to this paper (the link shows a file with slides of the project) i should just equalize the three channels obtaining some illuminant and device color invariant, but SIFT still fails!
