This page shows the full results for our feature description evaluation, for SIFT, our symmetry descriptors (SYMD), and a combination of the two descriptors (SIFT-SYM). For details on how we compute and match these descriptors between a pair of images, please see the paper. We show results in terms of precision-recall curves for matched features and their match scores (given the ground truth homography), created by varying the matching score threshold, for each pair of images in our dataset. We evaluate these three descriptors for four detectors: GRID (a synthetic set of perfectly matched detections, see paper for details), SIFT (i.e. DoG), and SYM-G and SYM-I (our two symmetry-based detectors).
The following table (Table 2 in the paper) reports the mean average precision for the different combinations of detector and descriptor on our dataset.
To see precision-recall curves for an image pair, just mouse over a thumbnail; to make the pair "stick", click on the icon. To see results for each detector, mouse over (or click) the GRID, SIFT, SYM-G, or SYM-I labels. The numbers in the legends indicate the average precision for each descriptor (higher is better).