SC-GAN: Subspace Clustering based GAN for Automatic Expression Manipulation

Highlights：

• We propose a subspace clustering based generative adversarial network (SC-GAN), which can simultaneously cluster multiple latent subspaces and generate diverse samples correspondingly. The SC-GAN can generate comparable facial attribute translation results with various styles of eyes and mouths.

• We carefully design and conduct experiments on various benchmarks to evaluate our SC-GAN. The experiments manifest that, our method can preserve more natural expression details and learn more information from the weakly-matched training dataset w.r.t. the state-of-the-art methods.

摘要

•We propose an unsupervised data-dividing method by integrating PCA-SIFT feature representation and SIFT K-means cluster. The method can split the dataset into smaller subspaces, wherein the distribution of the data can be learned more easily.•We propose a subspace clustering based generative adversarial network (SC-GAN), which can simultaneously cluster multiple latent subspaces and generate diverse samples correspondingly. The SC-GAN can generate comparable facial attribute translation results with various styles of eyes and mouths.•We carefully design and conduct experiments on various benchmarks to evaluate our SC-GAN. The experiments manifest that, our method can preserve more natural expression details and learn more information from the weakly-matched training dataset w.r.t. the state-of-the-art methods.

论文评审过程：Received 2 February 2021, Revised 27 May 2022, Accepted 21 September 2022, Available online 27 September 2022, Version of Record 13 October 2022.