Bridge the Intra-Class Gap: K-Shot Multi-Scale Intermediate Prototype Mining Transformer for Few-Shot Semantic...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot segmentation (FSS) aims to segment objects using limited support images.
• Existing methods struggle with intra-class diversity due to ignoring query-support category gaps.

Purpose of the Study:

• To propose a novel framework, KMIPMT, that addresses the category information gap in FSS.
• To enhance segmentation accuracy by capturing deterministic and adaptive knowledge at multiple scales.

Main Methods:

• Introduced intermediate prototypes to bridge the category information gap.
• Utilized a Transformer architecture (KMIPMT) for iterative feature enhancement.
• Propagated information from K-shot support and multi-scale query features to prototypes.

Main Results:

• Achieved state-of-the-art performance on PASCAL-5i, COCO-20i, and FSS-1000 benchmarks.
• Demonstrated remarkable performance gains with a simple yet effective approach.
• KMIPMT framework shows versatility across 3D point cloud, zero-shot, and weak-label FSS.

Conclusions:

• The proposed KMIPMT framework effectively enhances few-shot segmentation accuracy.
• Intermediate prototypes are crucial for handling intra-class diversity in FSS.
• KMIPMT offers a versatile and effective solution for various segmentation scenarios.