Dual Branch Multi-Level Semantic Learning for Few-Shot Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot semantic segmentation (FSS) segments novel objects using limited support examples.
• Existing prototype-based methods struggle with intra-class variations and inter-class similarities, leading to poor feature representations.
• Treating novel classes as background creates learning bias, hindering accurate foreground segmentation.

Purpose of the Study:

• To address the limitations of current FSS methods.
• To propose a novel dual-branch learning approach for improved segmentation performance.
• To enhance feature distinctiveness and generalizability to unseen classes.

Main Methods:

• A dual-branch learning framework combining class-specific and class-agnostic branches.
• Class-specific branch: Increases inter-class distance and decreases intra-class distance for better feature representation.
• Class-agnostic branch: Minimizes foreground feature distribution and maximizes foreground-background separation for generalizability; incorporates pixel-level and prototype-level semantic learning.

Main Results:

• The proposed method demonstrates effectiveness in few-shot semantic segmentation.
• Evaluated on PASCAL-5^i and COCO-20^i datasets across 1-shot and 5-shot settings.
• Achieved strong performance despite the method's simplicity.

Conclusions:

• The dual-branch learning method successfully overcomes feature representation challenges in FSS.
• It mitigates learning bias by effectively handling novel classes.
• The approach offers a simple yet effective solution for segmenting novel objects with limited data.