Exploring Spatial Diversity for Region-Based Active Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• State-of-the-art semantic segmentation relies on deep neural networks trained on large datasets.
• Pixel-level annotation for dense prediction tasks incurs significant costs.
• Region-based active learning offers a cost-effective alternative by selecting informative image regions for labeling.

Purpose of the Study:

• To reduce annotation costs for semantic segmentation while maintaining high performance.
• To investigate the benefits of local spatial diversity in region-based active learning.
• To develop a unified framework for region-based active learning incorporating spatial diversity and data uncertainty.

Main Methods:

• Proposed a unified optimization framework for region-based active learning.
• Incorporated local spatial diversity alongside traditional active selection criteria (e.g., data sample uncertainty).
• Applied the framework to Cityscapes and PASCAL VOC datasets for semantic segmentation.

Main Results:

• Enforcing local spatial diversity improved the performance of active learning methods.
• The proposed framework achieved 95% performance of fully supervised methods.
• The method utilized only 5-9% of labeled pixels, outperforming existing region-based active learning approaches.

Conclusions:

• Local spatial diversity is beneficial for region-based active learning in semantic segmentation.
• The developed framework significantly reduces annotation costs while preserving high accuracy.
• This approach offers a practical solution for large-scale semantic segmentation tasks with limited labeled data.