A Mutually Supervised Graph Attention Network for Few-Shot Segmentation: The Perspective of Fully Utilizing Limited...

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Fully supervised semantic segmentation demands extensive pixel-level annotations, which are time-consuming to acquire.
• Few-shot segmentation offers a solution by requiring minimal annotated samples for generalization to new categories.
• Optimizing the utilization of limited samples in few-shot segmentation remains a challenge.

Purpose of the Study:

• To propose a novel mutually supervised few-shot segmentation network.
• To enhance feature representation and preserve spatial information during segmentation.
• To establish a mutually supervised regime for improved segmentation accuracy.

Main Methods:

• Fusing feature maps from intermediate convolution layers for richer representations.
• Employing a graph attention network on bipartite graphs of support and query images.
• Utilizing attention maps for cross-modal enhancement in a mutual supervision framework.
• Integrating fused attention maps from intermediate layers into a graph reasoning layer.

Main Results:

• The proposed network effectively leverages limited annotated data.
• Experimental results on PASCAL VOC-5i and FSS-1000 datasets show superior performance.
• The method demonstrates significant improvements over baseline few-shot segmentation approaches.

Conclusions:

• The mutually supervised few-shot segmentation network achieves high effectiveness.
• The proposed approach addresses the challenge of limited sample utilization in segmentation.
• This method offers a promising direction for efficient and accurate few-shot image segmentation.