Active Learning under Simulated Class Imbalance and Label Shift in Binary Semantic Segmentation
The study simulated machine vision challenges—class imbalance and label shift—using two open-source datasets to test active learning strategies. It compared random sampling, entropy-based selection, and core-set selection against an imbalance-aware strategy. Findings show that entropy and core-set methods outperform random sampling, especially when data is highly imbalanced or the budget is moderate.