Uncertainty guided semi-supervised few-shot segmentation with prototype level fusion.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-Shot Semantic Segmentation (FSS) struggles with segmenting novel categories due to limited data and high intra-class variance.
• Transferring meta-knowledge to unseen categories is challenging, especially with diverse support-query pairs.

Purpose of the Study:

• To propose the Uncertainty Guided Adaptive Prototype Network (UGAPNet) for semi-supervised few-shot semantic segmentation.
• To address intra-class semantic bias by generating reliable pseudo-prototypes.
• To enhance generalized few-shot semantic segmentation by handling both seen and unseen classes.

Main Methods:

• Employs a shared meta-learner for pseudo-label prediction on unlabeled images.
• Incorporates an uncertainty estimation module to denoise pseudo-labels by quantifying prototype differences.
• Introduces a prototype rectification module for effective pseudo-prototypes and a generalized adaptive prototype.
• Proposes a Prototype-Level Fusion Strategy in prototypical contrastive space for generalized FSS.

Main Results:

• UGAPNet effectively generates reliable pseudo-prototypes, alleviating intra-class semantic bias.
• The uncertainty estimation and prototype rectification modules improve pseudo-label quality.
• The Prototype-Level Fusion Strategy successfully addresses confusion regions in generalized FSS.
• Experiments on benchmarks demonstrate significant effectiveness of UGAPNet and the fusion strategy.

Conclusions:

• UGAPNet offers a robust solution for semi-supervised few-shot semantic segmentation, particularly in challenging scenarios with high intra-class variance.
• The proposed methods enhance the ability to segment both seen and unseen classes simultaneously.
• The developed techniques show strong performance on benchmark datasets, advancing the field of FSS.