Quaternion-Valued Correlation Learning for Few-Shot Semantic Segmentation

Few-shot segmentation (FSS) aims to segment unseen classes given only a few annotated samples. Encouraging progress has been made for FSS by leveraging semantic features learned from base with sufficient training samples represent novel classes. The correlation-based methods lack the ability consider interaction of two subspace matching scores due inherent nature real-valued 2D convolutions. In this paper, we introduce quaternion perspective on correlation learning and propose Quaternion-valued Correlation Learning Network (QCLNet), aim alleviate computational burden high-dimensional tensor explore internal latent between query support images operations defined established algebra. Specifically, our QCLNet is formulated as hyper-complex valued network represents tensors in domain, which uses quaternion-valued convolution external relations when considering hidden relationship sub-dimension space. Extensive experiments PASCAL-5i COCO-20i datasets demonstrate that method outperforms existing state-of-the-art effectively. Our code available at https://github.com/zwzheng98/QCLNet