LOCMAX SIFT: Non-Statistical Dimension Reduction on Invariant Descriptors

The descriptors used for image indexing - e.g. Scale Invariant Feature Transform (SIFT) - are generally parameterized in very high dimensional spaces which guarantee the invariance on different light conditions, orientation and scale. The number of dimensions limit the performance of search techniques in terms of computational speed. That is why dimension reduction of descriptors is playing an important role in real life applications. In the paper we present a modified version of the most popular algorithm, SIFT. The motivation was to speed up searching on large feature databases in video surveillance systems. Our method is based on the standard SIFT algorithm using a structural property: the local maxima of these high dimensional descriptors. The weighted local positions will be aligned with a dynamic programming algorithm (DTW) and its error is calculated as a new kind of measure between descriptors. In our approach we do not use a training set, pre-computed statistics or any parameters when finding the matches, which is very important for an online video indexing application