|Comparison of Infomax and Maximum Variance Features|
One of the most popular feature extraction techniques (and one of the few that scales well enough to be feasible for problems involving large numbers of classes) is to select features that contain most of the signal energy, e.g. principal component analysis. The results below illustrate the improvements that can be obtained with a simple infomax-based feature selection algorithm. The algorithms, as well as the experimental set-up, are described in the CVPR 04 paper [ps, pdf]. The results presented here are from the Corel database, and were obtained for an image retrieval task. In all pictures, the image shown at the top-left was provided as a query and the remaining images are the top matches according to a retrieval system that relies on optimal feature selection. All system parameters are the same, except the algorithm used to determine the optimal features: infomax on the left, maximum variance on the right. Images which are retrieval errors (i.e. from a class different from that of the query) are indicated by red borders. Clicking on any image will show a higher resolution version, more suitable for detailed analysis of the retrieval performance.