3D Model Annotation based on Semi-Supervised Learning

† School of Computer and Information Technology, Northeast Petroleum University, Daqing, China Summary The purpose of annotation for 3D model is that it can automatically list the best suitable labels to describe the 3D models; it is an important part of the text-based 3D model retrieval. The existence of the semantic gap makes the result based on the similarity matching techniques needs to be improved. In order to improve the 3D model annotation performance using a large number of unlabeled samples, we propose a semi-supervised measure learning method to realize the 3D models multiple semantic annotation. A graph-based semi-supervised learning is firstly used to expand the training set, and the semantic words confidence of the models in the extension set is proposed. An improved relevant component analysis method is proposed in this paper to learn a distance measure based on the extended training set. Our approach is introduced to complete multiple semantic annotation task based on the learned distance measure. The test result on the PSB data set have shown that the method making use of the unlabeled samples has achieved a better annotation result when a small amount of labels were given.