AI Paper Atlas

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

ICCV 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Composed and text-based image retrieval · 8 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
fine-tune with angle-based entailment loss in hyperbolic space vs EuclideanCLIP ViT · HierOpenImages child-to-parenttop-5 accuracy: 75.63 -> 77.28Improved
fine-tune with entailment loss (Euclidean) vs pretrained CLIPCLIP ViT · HierOpenImages child-to-parenttop-5 accuracy: 53.04 -> 75.63Improved
add cross-image entailment samples (ablation)CLIP-hyp · HierOpenImages child-to-parenttop-5 accuracy: 73.56 -> 77.28Improved
add cross-image entailment samples (ablation)CLIP-hyp · HierOpenImages parent-to-childtop-5 accuracy: 71.87 -> 74.48Improved
add cross-image entailment samples (ablation)CLIP-hyp · HierOpenImages hierarchical retrievalOT distance top-150k: 15.98 -> 14.96Improved
use gated entailment angle metric instead of plain hyperbolic angle (ablation)CLIP-hyp · HierOpenImages child-to-parenttop-5 accuracy: 76.95 -> 77.28Improved
use gated metric instead of cosine similarity (ablation)MoCo-hyp · HierOpenImages parent-to-childtop-5 accuracy: 66.80 -> 69.50Improved
use hyperbolic entailment angle without gating (ablation)MoCo-hyp · HierOpenImages child-to-parenttop-5 accuracy: 68.62 -> 49.91Hurt
use hyperbolic instead of Euclidean fine-tuningMoCo-v2 · HierOpenImages recall top-150krecall: 74.84 -> 74.85No clear effect
use hyperbolic instead of Euclidean fine-tuningCLIP ViT · LVIS OT distance top-10kOT distance: 5.49 -> 5.56Hurt
use hyperbolic instead of Euclidean fine-tuningCLIP ViT · LVIS parent-to-childtop-5 accuracy: 28.84 -> 30.27Improved
fine-tune CLIP with hyperbolic entailmentCLIP ViT · zero-shot detection COCOaccuracy: 63.6 -> 73.9Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy