Dynamic-budget superpixel active learning for semantic segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Active learning reduces labeling costs in deep learning by prioritizing high-impact data points.
• Semantic segmentation tasks benefit from active learning, but static-budget strategies can lead to inefficient labeling.
• Variability in high-impact regions per image necessitates adaptive querying strategies.

Purpose of the Study:

• To present a novel dynamic-budget superpixel querying strategy for regional active learning in semantic segmentation.
• To enhance the querying efficiency and data efficiency of semantic segmentation workflows.
• To address the limitations of static-budget querying in active learning.

Main Methods:

• Developed a dynamic-budget superpixel querying strategy.
• Implemented and evaluated the strategy on two distinct datasets.
• Investigated the impact of superpixel size and budget scenarios (low and high).

Main Results:

• The dynamic-budget strategy proved more effective than static-budget querying across tested datasets.
• Achieved 5.6% mIoU improvement on an agriculture dataset and 2.4% mIoU on Cityscapes in a low-budget scenario.
• Demonstrated superior performance compared to static-budget querying at the same low total labeling budget.

Conclusions:

• The proposed dynamic-budget querying strategy is simple, effective, and improves data efficiency for semantic segmentation.
• This approach offers a more efficient alternative to static-budget querying in active learning.
• The strategy can be readily adapted to other regional active learning algorithms.