Research map / Label-efficient and robust learning / Domain adaptation and generalization

Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 7 citations (OpenAlex)

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 DADA, an adversarial domain adaptation framework with mixup-based data augmentation that aligns sample and proxy distributions to boost proxy-based deep metric learning losses (Proxy-Anchor, ProxyNCA++).

Problem

Proxy-based deep metric learning suffers from a large distribution gap between randomly initialized class proxies and image sample embeddings, which slows convergence and biases learning.

Approach

Treat samples and proxies as source and target domains, add an intermediate mixup domain of sample-proxy interpolations, and align them with a domain-level discriminator plus a category-level discriminator using nuclear-norm Wasserstein discrepancy, in a min-max adversarial game on top of existing proxy losses.

Key results

• PA+DADA (R50) reaches 72.9 R@1 on CUB200 and 92.1 R@1 on CARS196, improving over the PA baseline (R50) by 3.2pp and 4.4pp respectively.
• On In-Shop, PA+DADA achieves 93.0 R@1, the best among the compared methods in the table.
• On CUB200 ablation, the full DADA loss raises ProxyAnchor R@1 from 69.1 to 72.9 (+3.8pp) and ProxyNCA++ R@1 from 68.4 to 70.2 (+1.8pp).

Measured effects reported in the paper

Surprising findings

• Data augmentation alone (Aug without adversarial and classification losses) barely improves the baseline (+0.2pp R@1 on ProxyAnchor), and the mixed features only help when the adversarial losses are used.

Limitations

• Performance gains on SOP and In-Shop in high-recall metrics are limited, attributed to many classes with few samples each.
• The authors note limited learning efficiency on datasets with many classes and possible difficulty extending to continual learning.
• Vision Transformer backbones were not evaluated due to limited compute.

Experimental setup

Datasets: CUB-200-2011 CARS196 Stanford Online Products In-Shop Clothes Retrieval

Benchmarks: CUB-200-2011 CARS196 Stanford Online Products In-Shop Clothes Retrieval

Metrics: Recall@K MAP@R

Baselines: Proxy-Anchor ProxyNCA++ AVSL HIST IBC CRT DiVA S2SD DCML DRML ProxyGML PA+NIR

Builds on: Proxy-Anchor ProxyNCA++ Adversarial discriminative domain adaptation Nuclear-norm Wasserstein discrepancy mixup

Closest papers by meaning

• Non-Isotropy Regularization for Proxy-Based Deep Metric Learning (CVPR 2022 · similarity 0.89)
• Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.88)
• Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (ICLR 2024 · similarity 0.88)
• Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.88)
• Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.86)
• Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · similarity 0.85)