Analysis of Verification Methods for Indoor Image Matching

This paper reports on experiments for indoor image-based location recognition. The basic method makes use of three stages: visual bag-of-words for ranking, a voting method, and a final verification method, if the voting method does not produce a consensus. Such a tiered approach is necessary when there are several visually similar locations in the image database, such as often occurs in office buildings. Three experiments are reported here. In the first, three common term-weighting schemes are compared: ntf, ntfidf and BM25. Surprisingly ntf, the simplest scheme, is shown to be as accurate as BM25, and both are better than ntfidf. These results are surprising because BM25 has been experimentally shown to be one of the best weighting schemes for document information retrieval over many years, and ntfidf has been the preferred weighting scheme for visual BoW in most other image retrieval work. In the second experiment, two verification methods are compared: one based on the fundamental matrix, and one based on a simpler homography computation. Again, surprisingly, the simpler and more efficient homography based method is shown to perform as well as the fundamental matrix method despite the fact that the fundamental matrix method is more physically plausible. The overall system achieves a recognition rate of approximately 80% with a wrong match rate of only 2% (no decision on 18%) on a very challenging office building data set. In the third experiment, the system is evaluated on the same office building dataset with more than one query image. A significant improvement is observed in localisation performance and the overall system achieves a recognition rate of 96% with only two wrong image matches.