Completeness of Coding with Image Features
We investigated combinations of different feature detectors to get as much information as possible. Good results can be achieved with two or three detectors only in case they are highly complementary. Therefore we developed a measurement scheme for the completeness of a set of detectors with respect to the information contained in an image, which gives direct insight into the complementarity of detectors. It has been published in the International Journal of Computer Vision and in a shorter version on BMVC’09.
The completeness is expressed by the distance between two distributions over the image (top): A reference, represented by local entropy over scales (middle), and a feature coding density, represented by a Gaussian mixture distribution based on sets of features (bottom).
A presentation from my colleague Timo Dickscheid is available as a video on videolectures.net.
Title graphics: We applied our complementarity measure to multiple feature detectors and whole image datasets. Embedding all pairwise distances into a high-dimensional space and down-projecting them into three dimensions yields a map illustrating similar and complementary feature detectors.
