Cluster-Wise Ratio Tests for Fast Camera Localization

摘要: Feature point matching for camera localization suffers from scalability problems. Even when feature descriptors associated with 3D scene points are locally unique, as coverage grows, similar or repeated features become increasingly common. As a result, the standard distance ratio-test used to identify reliable image is overly restrictive and rejects many good candidate matches. We propose simple coarse-to-fine strategy that uses conservative approximations robust local ratio-tests can be computed efficiently using global approximate k-nearest neighbor search. treat these forward matches votes in pose space use them prioritize back-matching within clusters, exploiting co-visibility captured by model graph. This approach achieves state-of-the-art estimation results on variety of benchmarks, outperforming several methods more complicated data structures make assumptions pose. carry out diagnostic analyses difficult test dataset containing globally repetitive structure which suggest our successfully adapts challenges large-scale estimation.