A CPCA Principal Axes Corrected Approach Using the Direction of Maximum Normal Distribution

PCA-alignment only captures the relative location of 3D models' surface vertexes, and thus does not provide robust rotation normalization. In order to get more reasonable axes for the rotation normalization of 3D models, an approach to correct the principal axes by CPCA approach is proposed in this paper, using maximum normal distribution. The approach is called MNCPCA (maximum normal corrected PCA) for short. Two appearance's attributes which are vertexes position coordinates and normal directions are considered in the approach. To improve the robustness of the algorithm, firstly a group of referenced normals which are obtained by uniform sampling is used to compute normal distribution histogram, in order to avoid error in numerical calculation and incorrect normal; next normal distribution characteristic of 3D models is analyzed, judging whether the 3D model has an obvious maximum normal distribution by the normal distribution histogram. And then, based on the analysis, a strategy to correct CPCA principal axes is presented, only adjusting the principal axes for some 3D models which have obvious maximum normal distribution, and for the other models, the principal axes by CPCA is still unchanged. Experimental results show that the MNCPCA approach can get more reasonable axes than CPCA approach by human cognition.