For content-based 3D model retrieval, an improved depth image-based feature extraction algorithm is proposed. First, a 3-D model is preprocessed. Secondly, six depth images are generated in three principal directions in the normalized coordinate system. Thirdly, the eigenvectors of 3D model are obtained through 2D Fourier Transform of the depth images. Finally a new method is used for low-frequency sampling. Experiments show that the approach performed quite well despite its apparently simple approach. In our large 3D database, our approach is well for variant resolution models and holds satisfied computational costs.
Semantics-enhanced supervised deep autoencoder for depth image-based 3D model retrieval
