Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Disentangled Metric Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Disentangled Deep Metric Learning (DDML), a variational-information-bottleneck-based class-agnostic regularizer that disentangles class-specific and class-agnostic embedding features and plugs into existing proxy-based DML losses to improve generalization to unseen classes.
Problem
Proxy-based deep metric learning encourages class-specific features that generalize poorly to unseen classes, and NormSoftmax yields narrow, inconsistent uncertain regions between classes in the embedding space.
Approach
Extends the VIB objective to the open-set setting with an embedding bottleneck Z and a specific bottleneck Zs, adding agnostic loss (push class-agnostic features toward uniform class probabilities), specific loss, and a KL split loss to decouple features; this regularizer is attached to existing DML losses.
Key results
- DDML improves R@1 over NormSoftmax on CUB (64.65 to 65.96), CARS (83.54 to 84.85), and SOP (78.55 to 79.38) across three runs.
- DDML on Proxy Anchor with BN-Inception reaches 70.0 R@1 on CUB, 87.8 on CARS and 79.9 on SOP, versus 68.4, 86.1, 80.3 for the baseline in the main comparison table (SOP baseline value differs between tables).
- DDML on ResNet50 with Proxy Anchor reaches 71.5 R@1 on CUB, the best reported in Table 2 for that backbone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DDML to ProxyAnchor | ProxyAnchor · CUB | Recall@1: 68.36 ± 0.04 -> 69.91 ± 0.05 | Helped |
| Add DDML to ProxyAnchor | ProxyAnchor · CAR | Recall@1: 85.99 ± 0.07 -> 87.86 ± 0.06 | Helped |
| Add DDML to HypViT | HypViT · SOP | Recall@1: 85.86 ± 0.04 -> 86.08 ± 0.05 | Helped |
| Add agnostic loss (ablation) | NormSoftmax · CUB, CAR, SOP | Recall@1: 64.7 -> 64.9; 83.5 -> 84.2; 78.6 -> 78.8 | Helped |
| Add specific loss (ablation) | NormSoftmax · CUB, CAR | Recall@1: 64.7 -> 64.8; 83.5 -> 84.0 | Helped |
| Add specific loss (ablation) | NormSoftmax · SOP | Recall@1: 78.6 -> 78.6 | No clear effect |
| Add split loss (ablation) | NormSoftmax · CUB, CAR | Recall@1: 64.7 -> 64.5; 83.5 -> 83.1 | Hurt |
| Add split loss (ablation) | NormSoftmax · SOP | Recall@1: 78.6 -> 78.6 | No clear effect |
| Add agnostic and specific losses (ablation) | NormSoftmax · CUB, CAR, SOP | Recall@1: 64.7 -> 65.1; 83.5 -> 84.5; 78.6 -> 79.1 | Helped |
| Add agnostic and split losses (ablation) | NormSoftmax · CUB, CAR | Recall@1: 64.7 -> 64.7; 83.5 -> 83.4 | No clear effect |
| Add agnostic and split losses (ablation) | NormSoftmax · SOP | Recall@1: 78.6 -> 78.0 | Hurt |
| Add specific and split losses (ablation) | NormSoftmax · CUB, CAR, SOP | Recall@1: 64.7 -> 65.0; 83.5 -> 84.2; 78.6 -> 78.7 | Helped |
Surprising findings
- On SOP, the regularizer's gains are comparable to other regularization methods rather than clearly superior, with R@1 changes under 1%.
Limitations
- Gains are modest (typically under 1.5 R@1 points) and the SOP improvements are small.
- Performance may vary with number of classes and images per class, as the authors note for probabilistic models.
Experimental setup
Datasets: CUB-200-2011 CARS196 Stanford Online Products
Benchmarks: CUB-200-2011 CARS196 Stanford Online Products
Metrics: Recall@k
Baselines: NormSoftmax Proxy Anchor HypViT PS Metrix MS SoftTriple HIST HIER VPTSP-G DFML-PA
Builds on: variational information bottleneck NormSoftmax Proxy Anchor HypViT
Closest papers by meaning
- Distance Metric Learning with Joint Representation Diversification (ICML 2020 · similarity 0.89)
- Non-Isotropy Regularization for Proxy-Based Deep Metric Learning (CVPR 2022 · similarity 0.88)
- Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation (AAAI 2024 · similarity 0.88)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.86)
- Enriching Disentanglement: From Logical Definitions to Quantitative Metrics (NeurIPS 2024 · similarity 0.86)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in 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.