SIFT keypoint descriptors for range image analysis

This paper presents work in progress to extend the two-dimensional (2D) Scale Invariant Feature Transform (SIFT) into the 2.5 dimensional (2.5D) domain. Feature descriptors are extracted from range images of human faces and the form of these descriptors is analogous to the structure of Lowe’s 2D SIFT. Lowe’s descriptors are derived from the histogram of the image gradient orientations, computed over a Gaussian weighted local support region centred on each sampling (keypoint) location. We adapt this concept into the 2.5D domain by extracting the relative frequencies of the [-1,1] bounded range surface shape index and the relative frequencies of the range surface in-plane orientations simultaneously at each sampled keypoint location. Nine Gaussian weighted sub-regions, overlapped by one standard deviation, are used to sample each keypoint location and thereby construct a keypoint descriptor. Since these overlapped Gaussian sub-regions are spatially correlated, this sampling configuration reduces both the spatial aliasing and the sensitivity to small keypoint location errors in the extracted descriptors. Each histogram pair, extracted from each Gaussian weighted sub-region, is normalised and concatenated to form a feature descriptor that is tolerant to a degree of viewpoint rotational change. We have validated the current 2.5D SIFT formulation using synthetically rotated human face data over the range ±30° out-of-plane rotation, and demonstrate that our 2.5D SIFT maintains a similar matching performance to 2D SIFT applied to the (comparatively richer) intensity images of the same face