Content based 3D model Retrieval (CB3DR) is proved to be limited in performance due to the semantic gap between
low-level feature distance and high-level user intention. In order to capture semantics from models, we propose a new
framework which generates semantic subspaces for each category via corresponding variances of feature vectors. Then
vectorial and numerical semantic labels are composed from semantic subspaces. In the end, a Laplacian Eigenmaps
based manifold learning method is enhanced by these semantic labels and experiment results show an improvement in
performance with respect to classical Laplacian Eigenmaps method.