GSIFT: geometric scale invariant feature transform for terrain data

Abstract

In this work, we introduce GSIFT (Geometric Scale Invariant Terrain Feature Transform), geometric descriptors that are invariant to translation, rotation, and scaling. SIFT (Scale Invariant Feature Transform) descriptors have been found to be very successful in a variety of computer vision tasks. GSIFT expands SIFT by using the novel technique of binding features at different scales to associate priorities/importance to features based on its scale and persistence among different scales. As a first step to obtain scale invariance, we create a multi-scale pyramid for detecting important features such as maxima, minima, and saddle points for 1D and 2D height fields. We use relative height histograms as GSIFT with support regions determined by the priority of the feature. We use symmetric chi-square as the similarity measure to compare geometric features. Experiments with GSIFT and SIFT (on images constructed from height fields) on both synthetic and real height field data sets show that GSIFT provide comparable and in some cases, better results for data registration. GSIFT has the added advantage that it can be used by scientists in compressing and registering (or non-registering) height data, where it is important for them to understand which features to keep/discard or why the registration is good/poor, when the data is obtained at different scales from independent sources.