3D model classification using salient features for content representation

In 3D model retrieval area, many methods have been proposed. However, most of these methods are content-based, and focus on how to describe a model more accurately and how to measure the similarity between models. The most important problem in content-based retrieval is that there is a semantic gap between low-level visual features and high-level human semantic, so that most of the existing methods cannot tell the differences between a banana and a dolphin. How to bridge the semantic gap is becoming a new challenge. In this paper, we propose a hierarchical 3D model classification method to detect the semantics of 3D models. After getting the features of models, we use manifold learning methods to reduce the dimension of model features, and meanwhile get a semantic representation for the model. The reduced feature is called salient feature. Then, the salient features are used to train classifiers and the classifiers are used to detect model semantic. Our method is a semi-supervised learning method and can deal with the small sample problem. Experiments show that the method is promising.