We propose a novel feature representation approach based on a second-order configuration of local features to support the spatially stable matching of images. The approach is designed to be robust as regards anisotropic affine transformations without compromising its ability to reject mismatches. We revisit affine shape adaptation and extend its conclusion to characterize the second-order geometric coherence of local features. We also propose a second approach called Centrality-Sensitive Pyramid (CSP) for fast spatial neighborhood association. In an experiment with the Flickr Logos 32 dataset, our approach was able to achieve the highest accuracy yet reported for the retrieval protocol of this dataset, which was more than 8% higher than the second highest value.
Please click the thumbnail image to open the full-size PDF file.