Content-based 3D model retrieval considering the user’s relevance feedback

This dissertation deals with the problem of enabling precise and real-time retrieval of generic 3D models. We focus on three subproblems, namely, rotation normalization, discriminative shape representations and relevance feedback-based retrieval. Considering the first problem, we develop a rotation normalization method that computes the object’s principal axes using the orientation of the object’s surface. We also develop a hybrid rotation normalization scheme than combines both the surface distribution and the surface orientation distribution. Considering the second problem, we develop a set of 3D shape descriptors, namely, the Concrete Radialized Spherical Projection (CRSP), Hybrid and Panoramic Object Representation for Accurate Model Attributing (PANORAMA) shape descriptor. The CRSP descriptor is formed by computing a spherical function-based representation and using the spherical harmonic transform. The Hybrid descriptor is formed by combining the CRSP descriptor with a depth buffer-based representation. The PANORAMA descriptor is formed by projecting the surface of a 3D object and its orientation to the lateral surfaces of a set of cylinders that are centered at the object’s centroid and parallel to its principal axes. For each projection we compute the 2D Discrete Fourier Transform and Wavelet Transform. Considering the third problem, we develop a local relevance feedback technique which is based on shifting the feature vectors of 3D objects closer to their cluster centroid in 3D space.