Besides robustness and sparseness, the completeness of a set of local features wrt. the image content is important in many computer vision applications. We developed a practical tool for measuring such completeness.
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 first been published on BMVC’09 (find the paper here) and is described in more detail in the International Journal of Computer Vision (IJCV):
Dickscheid, Timo, Falko Schindler, and Wolfgang Förstner (2010): Coding Images with Local Features. International Journal of Computer Vision, doi:10.1007/s11263-010-0340-z, http://www.springerlink.com/index/10.1007/s11263-010-0340-z
The key idea is to express completeness by the distance between two distributions over the image (first image): A reference, represented by local entropy over scales (second image), and a feature coding density, represented by a Gaussian mixture distribution based on sets of features (third image). This way we intepret feature detection and description as image coding, in a similar manner as classical coding schemes like JPEG.
As a supplementary material, we provide here tables showing the average incompleteness score over image categories for all possible combinations of three detectors. Please refer to the original paper for a description of the applied detectors and image datasets.
If you want to compute completeness measures for your own detector (combinations), or on your own images, you can use our software for internal and research purposes. You find a zip of our main matlab functions here.