We present GraphMatch, an approximate yet efficient method for building thematching graph for large-scale structure-from-motion (SfM) pipelines. Unlikemodern SfM pipelines that use vocabulary (Voc.) trees to quickly build thematching graph and avoid a costly brute-force search of matching image pairs,GraphMatch does not require an expensive offline pre-processing phase toconstruct a Voc. tree. Instead, GraphMatch leverages two priors that canpredict which image pairs are likely to match, thereby making the matchingprocess for SfM much more efficient. The first is a score computed from thedistance between the Fisher vectors of any two images. The second prior isbased on the graph distance between vertices in the underlying matching graph.GraphMatch combines these two priors into an iterative "sample-and-propagate"scheme similar to the PatchMatch algorithm. Its sampling stage uses Fishersimilarity priors to guide the search for matching image pairs, while itspropagation stage explores neighbors of matched pairs to find new ones with ahigh image similarity score. Our experiments show that GraphMatch finds themost image pairs as compared to competing, approximate methods while at thesame time being the most efficient.
展开▼
