Deep Self-supervised Subspace Clustering with Triple Loss

Xiaotong Bu, Jiwen Dong, Mengjiao Zhang, Guang Feng, Xizhan Gao, Sijie Niu

Abstract: Deep subspace clustering (DSC) methods are widely used in various fields such as motion segmentation, image segmentation, and text mining. It uses the deep neural network to map high-dimensional features into low-dimensional latent subspace to achieve effective division of data. Nevertheless, DSC simply tends to learn representations based on auto-encoder, which can’t fully exploit the intrinsic structure of the data. In this paper, we design a novel approach called Deep self-supervised subspace clustering with triple data (DSSCT), which aims to uncover supervised information inherent in the data. Specifically, DSSCT leverages data augmentation and triple contrastive loss to obtain more effective low-dimensional representations that capture the similarity and difference among different samples. In addition, we introduce a dual self-expression matrix fusion strategy to further enhance the discriminant of the self-expression matrix used in DSSCT. To evaluate the performance of our proposed method, we conduct extensive experiments on several widely used public datasets and achieved excellent performance when compared with other state-of-art methods.