Research map / Label-efficient and robust learning / Recognition and long-tail
Non-Isotropy Regularization for Proxy-Based Deep 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 |
|---|---|---|---|
| Add NIR to ProxyAnchor | ProxyAnchor · CUB200-2011 | R@1: 64.4 $\pm$ 0.3 -> 66.0 $\pm$ 0.3 | Helped |
| Add NIR to ProxyAnchor | ProxyAnchor · CARS196 | R@1: 82.4 $\pm$ 0.4 -> 85.2 $\pm$ 0.3 | Helped |
| Add NIR to ProxyAnchor | ProxyAnchor · SOP | R@1: 78.0 $\pm$ 0.1 -> 78.9 $\pm$ 0.1 | Helped |
| Replace exponential loss scaling with SoftPlus or vary exponential temperature (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.3 $\pm$ 0.2; 66.1 $\pm$ 0.2; 65.8 $\pm$ 0.3; 85.1 $\pm$ 0.2; 85.0 $\pm$ 0.2; 84.8 $\pm$ 0.3 | No clear effect |
| Add gradient clipping (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 65.9 $\pm$ 0.1; 85.3 $\pm$ 0.1 | No clear effect |
| Stop NIR gradients from updating proxies (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.2 $\pm$ 0.4; 85.0 $\pm$ 0.2 | No clear effect |
| Include different-class pairs in NIR likelihood training (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 64.9 $\pm$ 0.3; 83.5 $\pm$ 0.4 | Hurt |
| Remove the proxy-based DML objective (ω = 0) (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.0 $\pm$ 0.3 -> 60.0 $\pm$ 0.4; 85.2 $\pm$ 0.2 -> 73.5 $\pm$ 0.6 | Hurt |
| Vary flow depth and subnet width (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.5 $\pm$ 0.5; 66.1 $\pm$ 0.3; 65.8 $\pm$ 0.4; 85.1 $\pm$ 0.4; 84.9 $\pm$ 0.2; 85.3 $\pm$ 0.1 | Mixed |
| Condition the flow on proxies at start, middle, or end only (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.4 $\pm$ 0.4; 66.1 $\pm$ 0.2; 65.7 $\pm$ 0.2; 85.1 $\pm$ 0.2; 84.9 $\pm$ 0.2; 84.8 $\pm$ 0.3 | No clear effect |
| Generate synthetic samples for proxy training (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.0% -> 64.8%; 85.2 $\pm$ 0.2 -> 84.6 $\pm$ 0.6 | Hurt |
| Use reverse distribution matching on generated samples (ablation) | ProxyAnchor + NIR · CUB200-2011 and CARS196 | R@1: 66.0% -> 63.2 $\pm$ 0.1; 85.2 $\pm$ 0.2 -> 83.4 $\pm$ 0.1 | Hurt |
Closest papers by meaning
- Proxy Synthesis: Learning with Synthetic Classes for Deep Metric Learning (AAAI 2021 · similarity 0.91)
- Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation (AAAI 2024 · similarity 0.89)
- Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.88)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.86)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (ICLR 2024 · similarity 0.85)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.85)
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.