Coverage-aware Semantic Representation Learning for Underrepresented Visual Concepts
Semantic Coverage Imbalance (SCI) is a bias where models ignore rare but important visual concepts due to long-tailed data. SemCovNet addresses this by learning to correct coverage disparities using three mechanisms: a Semantic Descriptor Map, Descriptor Attention Modulation, and a Descriptor–Visual Alignment loss. This framework reduces the Coverage Disparity Index (CDI), ensuring fair error rates across different semantic groups.