Research map / Detection and segmentation / Open-set and few-shot detection
Hyperbolic Contrastive Learning for Visual Representations Beyond Objects
Read the paperCheck 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 |
|---|---|---|---|
| add hyperbolic scene-object contrastive loss | MoCo-v2 pretrained on COCO or OpenImages · VOC, ImageNet-100, ImageNet-1k | linear-evaluation accuracy: 73.79 -> 75.40; 77.36 -> 79.08; 57.57 -> 58.51 | Helped |
| add hyperbolic scene-object contrastive loss | MoCo-v2, Dense-CL, and ORL · COCO detection and segmentation | APb and APm: 38.5 -> 40.6 APb; 34.8 -> 37.0 APm; 39.6 -> 42.5 APb; 35.7 -> 38.5 APm; 40.3 -> 41.4 APb; 36.3 -> 37.3 APm | Helped |
| use scene nodes rather than object nodes as the hierarchy center (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 76.96 | Hurt |
| replace hyperbolic distance with Euclidean similarity for scene-object pairs (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 76.68 | Hurt |
| set the hyperbolic scene-object loss weight to 0.01 instead of 0.1 (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 77.70 | Hurt |
| set the hyperbolic scene-object loss weight to 0.2 instead of 0.1 (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 78.64 | Hurt |
| increase the hyperbolic loss weight to 0.5 (ablation) | HCL with RSGD · ImageNet-100 | top-1 accuracy: 79.08 -> 0 | Hurt |
| replace RSGD with SGD (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 70.16 | Hurt |
| use a separate hyperbolic-branch encoder head instead of a shared head (ablation) | HCL · ImageNet-100 | top-1 accuracy: 79.08 -> 77.88 | Hurt |
| vary the Poincaré ball radius using the tested curvature settings (ablation) | HCL · ImageNet-1k | top-1 accuracy: 58.08; 58.31; 58.29; 58.51; 58.49 | Mixed |
| use the hyperbolic representation norm as the label-count ranking indicator | HCL representations · ImageNet-ReaL | NDCG: 0.633 -> 0.655 | Helped |
| use the hyperbolic representation norm as the label-count ranking indicator | HCL representations · COCO | NDCG: 0.791 -> 0.839 | Helped |
Closest papers by meaning
- Compositional Entailment Learning for Hyperbolic Vision-Language Models (ICLR 2025 · similarity 0.88)
- Learning Structured Representations with Hyperbolic Embeddings (NeurIPS 2024 · similarity 0.87)
- Unsupervised Feature Learning with Emergent Data-Driven Prototypicality (CVPR 2024 · similarity 0.87)
- Fully Hyperbolic Convolutional Neural Networks for Computer Vision (ICLR 2024 · similarity 0.87)
- Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping (CVPR 2024 · similarity 0.87)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · similarity 0.86)
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.