The article presents Hi-Mapper, a Visual Hierarchy Mapper that organizes visual scenes hierarchically using a novel probabilistic tree structure defined in hyperbolic space. This approach incorporates hierarchical relationships into the contrastive loss, allowing for efficient data utilization. Hi-Mapper's unique hierarchical decomposition and encoding facilitate enhanced global visual representation and significantly boost existing deep neural networks' capabilities. It has shown effective performance improvements in multiple dense prediction tasks such as image classification, object detection, and segmentation, marking a substantial advancement in visual scene analysis.
This paper introduces the Visual Hierarchy Mapper (Hi-Mapper), which utilizes a novel tree-like structure and a probabilistic approach in hyperbolic space for visual scene organization.
Through effective decomposition and encoding of visual hierarchies, Hi-Mapper significantly enhances the understanding and representation of complex scenes, leading to improved deep neural network (DNN) performance.
#visual-hierarchy-mapper #hyperbolic-geometry #deep-learning #image-classification #object-detection
Collection
[
|
...
]