Dual Prototypical Contrastive Learning for Few-shot Semantic Segmentation

TL;DR

This paper introduces a dual prototypical contrastive learning method for few-shot semantic segmentation, improving feature discrimination and generalization to unseen classes by using class-specific and class-agnostic contrastive losses.

Contribution

The novel dual contrastive learning framework enhances prototype discrimination and generalization in few-shot segmentation, outperforming existing methods on benchmark datasets.

Findings

Outperforms state-of-the-art on PASCAL-5i and COCO-20i datasets.

Effectively increases inter-class distance and reduces intra-class variance.

Enhances generalization to unseen classes through class-agnostic contrastive loss.

Abstract

We address the problem of few-shot semantic segmentation (FSS), which aims to segment novel class objects in a target image with a few annotated samples. Though recent advances have been made by incorporating prototype-based metric learning, existing methods still show limited performance under extreme intra-class object variations and semantically similar inter-class objects due to their poor feature representation. To tackle this problem, we propose a dual prototypical contrastive learning approach tailored to the FSS task to capture the representative semanticfeatures effectively. The main idea is to encourage the prototypes more discriminative by increasing inter-class distance while reducing intra-class distance in prototype feature space. To this end, we first present a class-specific contrastive loss with a dynamic prototype dictionary that stores the class-aware prototypes during training, thus enabling the same class prototypes similar and the different class prototypes to be dissimilar. Furthermore, we introduce a class-agnostic contrastive loss to enhance the generalization ability to unseen classes by compressing the feature distribution of semantic class within each episode. We demonstrate that the proposed dual prototypical contrastive learning approach outperforms state-of-the-art FSS methods on PASCAL-5i and COCO-20i datasets. The code is available at:https://github.com/kwonjunn01/DPCL1.