Multi Voxel Descriptor for 3D Texture Retrieval

In this paper, we present a new feature descriptors  which exploit voxels for 3D textured retrieval system when models vary either by geometric shape or texture or both. First, we perform pose normalisation to modify arbitrary 3D models  in order to have same orientation. We then map the structure of 3D models into voxels. This purposes to make all the 3D models have the same dimensions. Through this voxels, we can capture information from a number of ways.  First, we build biner voxel histogram and color voxel histogram.  Second, we compute distance from centre voxel into other voxels and generate histogram. Then we also compute fourier transform in spectral space.  For capturing texture feature, we apply voxel tetra pattern. Finally, we merge all features by linear combination. For experiment, we use standard evaluation measures such as Nearest Neighbor (NN), First Tier (FT), Second Tier (ST), Average Dynamic Recall (ADR). Dataset in SHREC 2014  and its evaluation program is used to verify the proposed method. Experiment result show that the proposed method  is more accurate when compared with some methods of state-of-the-art.