Beyond Support Samples: Incorporating Unlabeled Queries for Few-Shot Semantic Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot semantic segmentation (FSS) faces challenges with intra-class diversity due to limited annotated support data.
• Increasing annotated support images is impractical for few-shot learning frameworks.

Purpose of the Study:

• To enhance FSS accuracy by incorporating unlabeled query images for a more comprehensive category representation.
• To address the challenge of category bias in FSS using unlabeled data.

Main Methods:

• Proposed Unlabeled Query Integration Few-Shot Segmentation (UQI-FSS) framework.
• Developed an Unlabeled Query Integration Network (UQINet) to adaptively process unlabeled query images.
• Introduced Information Bridging, Query Fusion, and Adaptive Selection Modules within UQINet.

Main Results:

• UQINet adaptively extracts beneficial information and suppresses detrimental information from unlabeled query images.
• Significant performance improvements demonstrated over existing methods on FSS benchmarks.
• Versatility and practical value confirmed across four challenging segmentation scenarios.

Conclusions:

• UQI-FSS effectively leverages unlabeled query images to overcome limitations in few-shot semantic segmentation.
• The proposed UQINet architecture enhances segmentation accuracy and adaptability.
• This approach offers a practical solution for improving FSS in diverse applications.