Research map / Label-efficient and robust learning / Noisy labels

RoDA: Robust Domain Alignment for Cross-Domain Retrieval Against Label Noise

AAAI 2025 · Label-efficient and robust learning › Noisy labels · approach: Contrastive learning on noisy data · 5 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 RoDA, a framework for cross-domain image retrieval under noisy labels that combines selective division with relabeling/adaptive weighting (SDAL) and accumulative class-center domain alignment (ACCA), evaluated on Office-31, Office-Home, and Adaptiope.

Problem

Cross-domain image retrieval where both query and gallery domains contain label noise, which degrades supervised retrieval and hinders domain-invariant feature learning.

Approach

Divides data into clean and noisy subsets via kNN label inconsistency, relabels noisy samples with soft assignments to clean-data class prototypes, weights relabeled samples adaptively, and aligns query/gallery domains through class centers accumulated batch-by-batch in a domain loss.

Key results

• On Office-Home at 20% noise, RoDA reaches 57.25 best mAP versus 54.90 for JoCoR and 55.04 for clean-label CE-C.
• At 80% noise on Office-Home, RoDA achieves 36.99 mAP, well above the next best GCE at 25.36.
• RoDA's ablation shows the full combination (SD+AL+ACCA) reaches 49.47 best mAP on Office-Home 20% noise, versus 32.99 for CE only.

Measured effects reported in the paper

Surprising findings

• RoDA slightly underperforms Co-teaching at 20% noise on Adaptiope, which the authors attribute to category imbalance after division.
• RoDA at low noise surpasses cross entropy trained on clean labels (CE-C) in some settings, attributed to domain alignment.

Limitations

• Category imbalance after sample division can hurt performance on datasets with many categories (Adaptiope at 20% noise).
• Hyperparameters (beta, gamma, theta, alpha, lambda) are fixed and tuned per setup; sensitivity analysis is limited to beta and lambda.

Experimental setup

Datasets: Office-31 Office-Home Adaptiope

Benchmarks: Office-31 Office-Home Adaptiope

Metrics: mAP@All

Baselines: Cross Entropy Generalized Cross Entropy MAE Co-teaching JoCoR DivideMix NCE LongRemix DISC Robust-DivideMix

Builds on: DivideMix Generalized Cross Entropy Co-teaching JoCoR NCE

Closest papers by meaning

• DiDA: Disambiguated Domain Alignment for Cross-Domain Retrieval with Partial Labels (AAAI 2024 · similarity 0.91)
• Cross-Modal and Uni-Modal Soft-Label Alignment for Image-Text Retrieval (AAAI 2024 · similarity 0.88)
• Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.88)
• COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.86)
• Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport (AAAI 2024 · similarity 0.86)
• Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.86)