What If AI Understood Images Like We Do? This Model Might

A new Visual Hierarchy Mapper (Hi-Mapper) has been developed to analyze the hierarchical structure of visual scenes. The system uses a tree-like structure with probability distribution and learns hierarchical relationships in hyperbolic space. Hi-Mapper incorporates hierarchical interpretation into contrastive learning and efficiently identifies visual hierarchies. When integrated with existing deep neural networks, Hi-Mapper consistently improved performance across various image analysis tasks like classification, object detection, and segmentation.