Structure-Constrained Feature Extraction by Autoencoders for Subspace Clustering

Kewei Tang, Kaiqiang Xu, Zhixun Su, Wei Jiang, Xiaonan Luo, Xiyan Sun; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

Clustering the data points drawn from a union of subspaces, i.e., subspace clustering, is a hot topic in recent years. The assumption of the nonlinear subspace is more general for the real-world data, but also more difficult for the traditional methods. The deep neural network is a powerful technique extracting nonlinear features, so the autoencoders are usually adopted to handle this unsupervised problem. However, how to add constraints on the features to make them more effective is not heavily addressed by previous work. In this paper, we consider both the global and local structure of the features in the autoencoders by the low-rank property and Laplace operator, respectively. The low-rank property can make the learned features favoring similarity extraction by self-representation and the Laplace operator can help our method explore the useful information in the data set. Note that our way of placing constraints can also be employed in other deep neural networks. In fact, our method is closely associated with previous work. It can be viewed as the more general case of the structured autoencoders. Extensive experiments demonstrate the effectiveness of our method.