Visual word based location recognition in 3D models using distance augmented weighting
Friedrich Fraundorfer, Changchang Wu, Jan Frahm, Marc Pollefeys · 2008
For visual word based location recognition in 3D models we propose a novel distance-weighted scoring scheme. Matching visual words are not treated as perfect matches anymore but are weighted with the distance of the original SIFT feature vectors before quantization. To maintain the scalability and efficiency of vocabulary tree based approaches PCA compressed SIFT feature vectors are used instead of the original SIFT features. A different eigenspace is computed for each vocabulary tree cell to benefit from the variance reduction as result of the partitioned SIFT feature space. Experiments show a significant improvement in retrieval quality by incorporating the distance with small costs in computational time and memory. 1