Beyond Flat Labels: Level-Restricted Contrastive Learning for Hierarchical Fine-Grained Vision Classification
Standard contrastive learning fails with hierarchical labels because it assumes a flat label space, causing inconsistent predictions across different taxonomic levels. This work introduces a level-restricted framework that restricts comparisons to single taxonomic levels and uses group-balanced supervision across all levels. This approach significantly improves both classification accuracy and the model's ability to maintain correct hierarchical structure in its learned representations.