ARNet: Active-Reference Network for Few-Shot Image Semantic Segmentation

Abstract

Few-shot segmentation has become a research focus for making predictions on unseen classes. However, most methods rely on pixel-level annotation, requiring significant manual effort, and the diversity in feature representation of same-category objects due to size, appearance, or layout differences poses additional challenges. To address these issues, the proposed active-reference network (ARNet) introduces an active-reference mechanism that supports accurate segmentation with cooccurrent objects in support or query images and relaxes the need for precise pixel-level labeling by allowing weak boundary labeling. Additionally, a category-modulation module (CMM) is applied to fuse features from multiple support images, selectively forgetting irrelevant information and enhancing key features. Experiments on the PASCAL-5i dataset show ARNet achieves a mean IOU score of 56.5% for 1-shot and 59.8% for 5-shot segmentation, outperforming the current state-of-the-art methods by 0.5% and 1.3%, respectively.

Yirui Wu

Young Professor, CCF Senior Member

My research interests include Computer Vision, Artifical Intelligence, Multimedia Computing and Intelligent Water Conservancy.