Research map / Label-efficient and robust learning / Recognition and long-tail
Hyperbolic Vision Transformers: Combining Improvements in Metric Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace spherical embeddings and cosine distance with hyperbolic embeddings and hyperbolic distance | ViT metric-learning models · CUB-200-2011, Cars-196, SOP, In-Shop | Recall@1: CUB-200-2011: 83.2 -> 84.0; Cars-196: 78.5 -> 82.7; SOP: 82.5 -> 85.5; In-Shop: 90.8 -> 92.7 | Helped |
| replace spherical embeddings and cosine distance with hyperbolic embeddings and hyperbolic distance | DINO metric-learning models · CUB-200-2011, Cars-196, SOP, In-Shop | Recall@1: CUB-200-2011: 76.0 -> 78.3; Cars-196: 81.9 -> 86.0; SOP: 82.0 -> 84.6; In-Shop: 90.4 -> 92.6 | Helped |
| use 8 × 8 rather than 16 × 16 image patches (ablation) | Hyp-DINO · Cars-196 | Recall@1: 86.0 -> 90.4; +4.4% | Helped |
| use 8 × 8 rather than 16 × 16 image patches (ablation) | Hyp-DINO, 384-dimensional embeddings · Cars-196 | Recall@1: 89.2 -> 92.8 | Helped |
| set curvature to 0.01, 0.05, or 0.3 instead of default 0.1 (ablation) | Hyp-ViT · In-Shop | Recall@1: Default: 92.4; c = 0.01: 92.3; c = 0.05: 92.4; c = 0.3: 92.3 | No clear effect |
| increase curvature to 0.5 or 1.0 instead of default 0.1 (ablation) | Hyp-ViT · In-Shop | Recall@1: Default: 92.4; c = 0.5: 91.8; c = 1.0: 90.0 | Hurt |
| reduce head dimensionality to 16, 32, or 64 instead of 128 (ablation) | Hyp-ViT · In-Shop | Recall@1: Default: 92.4; dim. 16: 88.6; dim. 32: 90.2; dim. 64: 91.6 | Hurt |
| reduce batch size to 200 (ablation) | Hyp-ViT · In-Shop | Recall@1: Default: 92.4; batch size 200: 92.0 | Hurt |
| use batch size 400 or 1600 instead of default 900 (ablation) | Hyp-ViT · In-Shop | Recall@1: Default: 92.4; batch size 400: 92.5; batch size 1600: 92.4 | No clear effect |
| use ImageNet-21k-pretrained ViT-S instead of DINO-pretrained ViT-S | Hyp metric learning · CUB-200-2011 | Recall@1: 78.3 -> 84.0 | Helped |
| use DINO-pretrained ViT-S instead of ImageNet-21k-pretrained ViT-S | Hyp metric learning · Cars-196 | Recall@1: 82.7 -> 86.0 | Helped |
Closest papers by meaning
- Unsupervised Hyperbolic Metric Learning (CVPR 2021 · similarity 0.91)
- Hyperbolic Image Embeddings (CVPR 2020 · similarity 0.91)
- Klein Hyperbolic Metric Learning (ICML 2026 · similarity 0.90)
- Robust Hyperbolic Learning with Curvature-Aware Optimization (NeurIPS 2025 · similarity 0.89)
- Hyperbolic Embeddings of Supervised Models (NeurIPS 2024 · similarity 0.89)
- Fully Hyperbolic Convolutional Neural Networks for Computer Vision (ICLR 2024 · similarity 0.89)
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.