Research map / Label-efficient and robust learning / Noisy labels
Dual Self-Paced Cross-Modal Hashing
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 DSCMH, a cross-modal hashing method that applies self-paced learning at both instance and feature levels to learn robust hash codes from easy to hard samples.
Problem
Cross-modal hashing methods treat all instances and features equally, making them vulnerable to noise and outliers in multimodal data.
Approach
Assigns learnable reliability weights to each instance and each feature via a self-paced regularizer, gradually increasing the weights during training so the model learns from easy to hard in both levels, combined with cross-space consensus learning between latent space and Hamming codes.
Key results
- DSCMH achieves 78.76% mAP for I→T on MIRFlickr at 64 bits, versus 76.96% for the best baseline AMSH.
- On MIRFlickr with injected noise (T→I, 8 bits), DSCMH reaches 80.00% mAP versus 79.08%-80.00% range of baselines, outperforming AMSH's 78.83%.
- Ablation on MIRFlickr shows removing dual self-paced learning (DSCMH-2) drops I→T 64-bit mAP from 78.76% to 76.14%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DSCMH instead of AMSH | Cross-modal hashing · MIRFlickr, I → T, 64 bits | mAP (%): 76.96 -> 78.76 | Helped |
| Use DSCMH instead of AMSH | Cross-modal hashing · IAPR-TC12, I → T, 64 bits | mAP (%): 53.66 -> 56.75 | Helped |
| Use DSCMH instead of FDDH | Cross-modal hashing · NUS-WIDE, I → T, 64 bits | mAP (%): 68.60 -> 69.26 | Helped |
| Use DSCMH instead of AMSH | Cross-modal hashing · MIRFlickr, T → I, 64 bits | mAP (%): 83.07 -> 83.59 | Helped |
| Use DSCMH instead of AMSH | Cross-modal hashing · IAPR-TC12, T → I, 64 bits | mAP (%): 66.32 -> 67.21 | Helped |
| Use DSCMH instead of FDDH | Cross-modal hashing · NUS-WIDE, T → I, 64 bits | mAP (%): 81.58 -> 80.99 | Hurt |
| Use DSCMH instead of FDDH | Cross-modal hashing · MIRFlickr with noise, I → T, 64 bits | mAP (%): 63.75 -> 66.76 | Helped |
| Use DSCMH instead of AMSH | Cross-modal hashing · MIRFlickr with noise, T → I, 64 bits | mAP (%): 82.34 -> 82.90 | Helped |
| Replace consensus latent representations with specific-view latent representations (ablation) | DSCMH · MIRFlickr, I → T, 64 bits | mAP (%): 78.76 -> 78.24 | Hurt |
| Replace consensus latent representations with specific-view latent representations (ablation) | DSCMH · MIRFlickr, T → I, 64 bits | mAP (%): 83.59 -> 82.34 | Hurt |
| Add dual self-paced learning (ablation) | DSCMH · MIRFlickr, I → T, 64 bits | mAP (%): 76.14 -> 78.76 | Helped |
| Add dual self-paced learning (ablation) | DSCMH · MIRFlickr, T → I, 64 bits | mAP (%): 81.86 -> 83.59 | Helped |
Limitations
- Relies on RBF kernel features with randomly chosen anchors and hyperparameters tuned per dataset.
- MTFH could not run on NUS-WIDE due to memory limits of its n x n similarity matrix.
Experimental setup
Datasets: MIRFlickr IAPR-TC12 NUS-WIDE
Benchmarks: MIRFlickr IAPR-TC12 NUS-WIDE
Metrics: mAP precision-recall curve training time
Baselines: RFDH LCMFH DLFH MTFH FCMH FDDH BATCH EDMH DAH ALECH WASH AMSH
Builds on: self-paced learning LCMFH AMSH
Closest papers by meaning
- Robust Self-Paced Hashing for Cross-Modal Retrieval with Noisy Labels (AAAI 2025 · similarity 0.95)
- Robust Self-reflective Hashing for Cross-modal Retrieval with Noisy Label (ICML 2026 · similarity 0.92)
- Meta-Guided Sample Reweighting for Robust Cross-Modal Hashing Retrieval with Noisy Labels (AAAI 2026 · similarity 0.91)
- Deep Cross-Modal Hashing (CVPR 2017 · similarity 0.91)
- Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (AAAI 2025 · similarity 0.89)
- Semi-Supervised Online Cross-Modal Hashing (AAAI 2025 · similarity 0.89)
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.