HierVision: Standardized and Reproducible Hierarchical Sources for Vision Datasets

Abstract

Many vision tasks benefit from hierarchical label structures — such as class taxonomies from WordNet — but sourcing, cleaning, and validating these hierarchies is done inconsistently across works, making fair comparison difficult. HierVision provides a standardized, reproducible toolkit for deriving hierarchical label sources for common vision benchmarks (ImageNet, iNaturalist, and others). By unifying how hierarchies are constructed and evaluated, HierVision enables fair comparison of hierarchy-aware learning methods and lowers the barrier for researchers adopting structured label supervision.

BibTeX

@inproceedings{kasarla2025hiervision,
  title={HierVision: Standardized and Reproducible Hierarchical Sources for Vision Datasets},
  author={Kasarla, Tejaswi and Rooparaghunath, Ruthu Hulikal and D'Arrigo, Stefano and Mago, Gowreesh and Jha, Abhishek and Ayoughi, Melika and Mishra, Swasti Shreya and Manzano Rodr{\'i}guez, Ana and Long, Teng and Ghadimi Atigh, Mina and van Spengler, Max and Mettes, Pascal},
  booktitle={ECCV Workshop on Beyond Euclidean},
  year={2025}
}