K. Grauman and T. Darrell.  The Pyramid Match Kernel: Discriminative Classification with Sets of Image Features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Beijing, China, October 2005.

Updated Pyramid Match Kernel results on the Caltech101 data set

The plot below shows (in blue) the current object recognition results using the pyramid match kernel (PMK) on the Caltech101 data set, compared against other published results. The PMK scores shown have been normalized according to the number of novel test images per class; that is, the mean recognition rate per class is the average normalized score for all 101 categories. Recognition accuracy is plotted for varying numbers of training examples per class, specifically [1,3,5,10,15,20,25,30]. For each training set size, we are displaying the pyramid match accuracy averaged over 10 runs, where for each run we randomly select the training examples and use all the remaining database images as test examples.

A standard number of training examples per class is 15. For this size of training set, we obtain a recognition performance of 50% averaged over all classes, where again the number of correct predictions per class has been normalized by the number of examples in that class.

This updates the results in our ICCV 2005 paper, in which an accuracy number is only provided for one training set size, and the performance was not normalized per class.

For these results, the pyramid match operated on sets of SIFT features projected to 10 dimensions using PCA, with each appearance descriptor concatenated with its corresponding positional feature. The features were extracted on a uniform grid from the images (i.e., no interest operator was applied). Classification was done with a one-vs-all support vector machine (SVM).

The computation time for a single pyramid match kernel value averaged 0.002 seconds, and the sets had on average 1140 features each.

Pyramid match kernel reference:

K. Grauman and T. Darrell.  The Pyramid Match Kernel: Discriminative Classification with Sets of Image Features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Beijing, China, October 2005.

Timeline of results as of CVPR 2006: