Automatic Extraction Of Invariant Features For Object Recognition

A powerful technique for three-dimensional object recognition has been the use of geometric invariants: measurable relationships between geometric objects that are invariant to transformations such as projection. Because of the invariance, these measurements will be the same whether measured on the actual three-dimensional object, or in the image. Therefore, objects in the image can be recognized if the same invariant can be found. In this paper, we investigate the automatic extraction of cross-ratio invariant features for object recognition. We show that clustering is a promising technique in this extraction, because it reduces the dependency on tuning parameters in the image processing phase.