Research map / Vision-language and multimodal / Human and embodied modeling
Learning Visual Hierarchies in Hyperbolic Space for Image Retrieval
Ziwei Wang, Sameera Ramasinghe, Chenchen Xu, Julien Monteil, Loris Bazzani, Thalaiyasingam Ajanthan
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with angle-based entailment loss in hyperbolic space vs Euclidean | CLIP ViT · HierOpenImages child-to-parent | top-5 accuracy: 75.63 -> 77.28 | Improved |
| fine-tune with entailment loss (Euclidean) vs pretrained CLIP | CLIP ViT · HierOpenImages child-to-parent | top-5 accuracy: 53.04 -> 75.63 | Improved |
| add cross-image entailment samples (ablation) | CLIP-hyp · HierOpenImages child-to-parent | top-5 accuracy: 73.56 -> 77.28 | Improved |
| add cross-image entailment samples (ablation) | CLIP-hyp · HierOpenImages parent-to-child | top-5 accuracy: 71.87 -> 74.48 | Improved |
| add cross-image entailment samples (ablation) | CLIP-hyp · HierOpenImages hierarchical retrieval | OT distance top-150k: 15.98 -> 14.96 | Improved |
| use gated entailment angle metric instead of plain hyperbolic angle (ablation) | CLIP-hyp · HierOpenImages child-to-parent | top-5 accuracy: 76.95 -> 77.28 | Improved |
| use gated metric instead of cosine similarity (ablation) | MoCo-hyp · HierOpenImages parent-to-child | top-5 accuracy: 66.80 -> 69.50 | Improved |
| use hyperbolic entailment angle without gating (ablation) | MoCo-hyp · HierOpenImages child-to-parent | top-5 accuracy: 68.62 -> 49.91 | Hurt |
| use hyperbolic instead of Euclidean fine-tuning | MoCo-v2 · HierOpenImages recall top-150k | recall: 74.84 -> 74.85 | No clear effect |
| use hyperbolic instead of Euclidean fine-tuning | CLIP ViT · LVIS OT distance top-10k | OT distance: 5.49 -> 5.56 | Hurt |
| use hyperbolic instead of Euclidean fine-tuning | CLIP ViT · LVIS parent-to-child | top-5 accuracy: 28.84 -> 30.27 | Improved |
| fine-tune CLIP with hyperbolic entailment | CLIP ViT · zero-shot detection COCO | accuracy: 63.6 -> 73.9 | Improved |
Closest papers by meaning
- HiHPQ: Hierarchical Hyperbolic Product Quantization for Unsupervised Image Retrieval (AAAI 2024 · similarity 0.88)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.88)
- Compositional Entailment Learning for Hyperbolic Vision-Language Models (ICLR 2025 · similarity 0.87)
- Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping (CVPR 2024 · similarity 0.87)
- Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval (AAAI 2024 · similarity 0.87)
- Learning Structured Representations with Hyperbolic Embeddings (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
