Multi-view Latent Space Learning Based on Local Discriminant Embedding

Abstract

In many computer vision systems, one object can be described by multi-view data. Compared with individual view, multi-view data can contain complete and complementary information of the problem. But when views capture information which is uniquely but not complete enough to give an uniform learning performance, multi-view data may degrade the learning performance and it is therefore not an ideal solution to simply concatenate multiple views into single view. In this paper, we proposed an multi-view latent space learning algorithm which assume that multi-view data is extracted from the same latent space via distinct transformation. Under this assumption, our algorithm can have a good performance even though views are not complete and the space obtained can contain the valuable information of each view as well as get the underlying connections between multi-view data. Due to the local discriminant embedding of the input space, this multi-view latent space is more suitable for classification or recognition problems. The proposed algorithm is evaluated on two tasks: indoor scene classification and abnormal objects classification on MIT scene 67, Abnormal Objects database respectively. Extensive experiments show that the algorithm we proposed achieves comparable improvements when compared with many other outstanding methods.