WCSE

Learning to Cluster Faces via Hypergraph Convolution with Transformer on Large Graph

Abstract—Face clustering is a very useful method in image annotation, image retrieval and other real-world applications. The main challenge is that with the increasing data scale, the large graph constructed by KNN is difficult to train due to out of memory. At the same time, the image features of the previous face clustering methods are extracted through CNN, which is not conducive to obtaining the global features of the image. In this paper, we propose a method that uses transformer to extract image features. Then, we use the extracted image features to construct a large graph by KNN, and we randomly partition the large graph into several non-overlapping subgraphs through METIS. In addition, we use hypergraph convolution to learn deeper high-order graph structure data. Experiments on both MS-Celeb-1M and DeepFashion show that our method achieves state-of-the-art performance, eg.90.08% in pair-wise F-score on MS-Celeb-1M.

Index Terms—Face clustering, Transformer, Hypergraph convolution network, Large graph

Dengdi Sun
School of Artificial Intelligence, Anhui University, CHINA
Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, CHINA
Zhizhong Huang, Bin Luo
School of Computer Science and Technology, Anhui University, CHINA
Zhuanlian Ding
Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, CHINA

[Download]

Cite: Dengdi Sun, Zhizhong Huang, Bin Luo, Zhuanlian Ding, "Learning to Cluster Faces via Hypergraph Convolution with Transformer on Large Graph" Proceedings of 2023 the 13th International Workshop on Computer Science and Engineering (WCSE 2023), pp. 118-128, June 16-18, 2023.