Interpolation of Novel Object Views from Sample Views

In this article we address the problem of threedimensional object recognition from two-dimensional views. We use a viewer-centered model of object representation and interpolate novel views from stored sample views. The sample views are represented by graphs which are labeled with Gabor wavelet responses as local descriptors of object points. The positions of the object points in a novel view are linear combinations of the corresponding point positions in the sample views, and the novel feature vectors are linear combinations of the Gabor responses in the sample views. From such an interpolated graph we reconstruct the novel view and analyse its quality for different poses of the novel view in relation to the sample views. Within the covered range of about 30◦ tilt and 40◦ pan viewing angle between the sample views we obtain good interpolation qualities. This leads to a number of about 36 views which is sufficient to represent the upper viewing hemisphere of an arbitrary object. Our results are consistent with current findings of biological and psychological research. In addition, the idea of the proposed algorithm is suitable to be applied in data compression.