Abstract

The purpose of the few-shot segmentation task is to segment images containing new categories using only a few labeled samples. Existing methods typically extract features for support and query branches through a Siamese encoder and learn category-related information from the labeled support image to guide the segmentation of the query image. However, the extracted features are usually low- or middle-resolution, and these methods only exploit marginal distributions and neglect joint distributions. To address these issues, a High-Resolution representation and Brownian distance covariance learning (HRB) method is proposed for the task. Firstly, a high-resolution Siamese encoder is adopted to extract high-to-low-resolution features for two branches. Then, a pyramid feature joint module is proposed to learn high-resolution feature representations, and meanwhile, a new cross-excitation module is designed to enhance their common semantic information. Furthermore, a new similarity metric module is developed, i.e., the Brownian distance covariance metric, which estimates the semantic mapping relationship between the joint distribution features of the two branches. Extensive experiments on three benchmark datasets (PASCAL-5ⁱ, FSS-1000, and COCO-20ⁱ) prove that the proposed HRB attains state-of-the-art performance. The code is available at https://github.com/Saralyliu/HRB.

中文翻译：

---

基于高分辨率表示和布朗距离协方差学习的少镜头分割

摘要

少样本分割任务的目的是仅使用少量标记样本来分割包含新类别的图像。现有方法通常通过连体编码器提取支持和查询分支的特征，并从标记的支持图像中学习类别相关信息以指导查询图像的分割。然而，提取的特征通常是低或中分辨率的，并且这些方法仅利用边缘分布而忽略联合分布。为了解决这些问题，针对该任务提出了高分辨率表示和布朗距离协方差学习（HRB）方法。首先，采用高分辨率暹罗编码器来提取两个分支的高分辨率到低分辨率的特征。然后，提出了金字塔特征联合模块来学习高分辨率特征表示，同时设计了新的交叉激励模块来增强它们的共同语义信息。此外，还开发了一种新的相似性度量模块，即布朗距离协方差度量，它估计两个分支的联合分布特征之间的语义映射关系。对三个基准数据集（PASCAL-5 ⁱ、FSS-1000 和 COCO-20 ⁱ）的广泛实验证明，所提出的 HRB 具有最先进的性能。代码可在 https://github.com/Saralyliu/HRB 获取。