Transductive meta-learning with enhanced feature ensemble for few-shot semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Few-shot semantic segmentation requires accurate pixel-level classification with limited labeled data.
• Existing methods struggle with feature diversity and false positives from base classes.

Purpose of the Study:

• To propose a novel transductive end-to-end method for few-shot semantic segmentation.
• To enhance feature representation and mitigate false positives for improved segmentation accuracy.

Main Methods:

• Ensemble of visual features from pretrained classification and semantic segmentation networks.
• Utilizing a pretrained semantic segmentation network as a base class extractor.
• A two-step segmentation approach with transductive meta-learning for intra-class and intra-object similarity learning.

Main Results:

• Achieved state-of-the-art performance on Pascal VOC and COCO datasets for 1-shot and 5-shot segmentation.
• Demonstrated significant improvements with minimal trainable parameters (2.98M).
• Effectively mitigated false positives and improved segmentation accuracy.

Conclusions:

• The proposed method offers a robust solution for few-shot semantic segmentation.
• Ensemble features and transductive meta-learning are effective in improving segmentation performance.
• The approach provides a strong baseline with high efficiency and accuracy.