3D-DFT Spectrum and Cepstrum of Dense Local Cubes for 3D Model Retrieval

Traditional 3D-DFT features used for 3D model retrieval often extracted the spectral/cepstral features from the spectrum/cepstrum of the holistic 3D model. These approaches seldom consider the local properties, which may reflect to some extent the shape properties of variant parts of a 3D model. In our prior work, we extracted some local 3D-DFT spectral and cepstral descriptors from each local octant of a 3D model. However, these eight octants can only characterize some specific region of a 3D model, never covering any possible regions. To this end, we will decompose a 3D model into dense local cubes, with these cubes cover local regions of a 3D model as more as possible. 3D-DFT spectral descriptors and cepstral descriptors are then extracted from each local cube and combined to form the local descriptor for similarity matching. Experiments conducted on the Princeton Shape Benchmark (PSB) database have shown that the proposed dense local 3D-DFT descriptors based on dense local cube decomposition outperform other 3D-DFT descriptors as well as octant-based local descriptors.