DRNet: Double Recalibration Network for Few-Shot Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot segmentation aims to segment images using minimal annotated examples.
• Existing methods struggle with high intra-class variance between query and support images.
• Robust feature embedding is crucial for accurate segmentation in limited-data scenarios.

Purpose of the Study:

• To enhance the robustness of few-shot segmentation models against intra-class variance.
• To introduce a novel network architecture, the Double Recalibration Network (DRNet), for improved segmentation.
• To leverage semantic-aware knowledge within the query image for self-adaptation in segmentation.

Main Methods:

• Proposed the Double Recalibration Network (DRNet) incorporating Self-adapted Recalibration (SR) and Cross-attended Recalibration (CR) modules.
• SR module utilizes query image semantics to refine initial object region prediction.
• CR module refines query feature representation by cross-attending to support image foreground features.

Main Results:

• DRNet achieved state-of-the-art performance on PASCAL-5^i and COCO-20^i benchmarks.
• Achieved mIoU of 63.6% (1-shot) and 64.9% (5-shot) on PASCAL-5^i.
• Achieved mIoU of 44.7% (1-shot) and 49.6% (5-shot) on COCO-20^i.

Conclusions:

• DRNet effectively addresses intra-class variance in few-shot segmentation.
• The proposed recalibration modules significantly improve the accuracy of target region mining.
• DRNet offers a robust and effective solution for challenging few-shot segmentation tasks.