Research map / Label-efficient and robust learning / Noisy labels
Distribution-Consistency-Guided Multi-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 DCGMH, a supervised multi-modal hashing method that filters and corrects noisy labels using a discovered consistency between label 1-0 patterns and similarity-score distributions to category centers.
Problem
Supervised multi-modal hashing assumes clean training labels, but manual annotation errors degrade retrieval performance under noisy labels.
Approach
Discovers that label presence/absence matches high-low similarity between hash codes and learnable category centers; uses this consistency to split training data into clean and noisy sets, corrects high-confidence noisy labels via nearest clean neighbors, and treats low-confidence ones as unlabeled for contrastive learning.
Key results
- At 40% noisy labels, DCGMH reaches 0.850 MAP on MIR Flickr at 128 bits versus 0.806 for the best baseline SAPMH.
- On NUS-WIDE at 64 bits, DCGMH achieves 0.755 MAP compared to 0.737 for DIOR.
- Ablation shows removing noise filtering (DCGMH-I) drops MIR Flickr 64-bit MAP from 0.846 to 0.784.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DCGMH instead of DIOR | multi-modal hashing · MIR Flickr, 40% noisy labels, 64 bits | MAP: 0.820 -> 0.846 | Helped |
| apply DCGMH instead of DIOR | multi-modal hashing · NUS-WIDE, 40% noisy labels, 64 bits | MAP: 0.737 -> 0.755 | Helped |
| apply DCGMH instead of STBMH | multi-modal hashing · MS COCO, 40% noisy labels, 64 bits | MAP: 0.564 -> 0.631 | Helped |
| add label filtering and reconstruction rather than training directly on original noisy data (ablation) | DCGMH · MIR Flickr, NUS-WIDE, and MS COCO, 40% noisy labels, 64 bits | MAP: 0.784 -> 0.846; 0.721 -> 0.755; 0.610 -> 0.631 | Helped |
| add correction of high-confidence noisy labels rather than treating all noisy labels as unlabeled (ablation) | DCGMH · MIR Flickr, NUS-WIDE, and MS COCO, 40% noisy labels, 64 bits | MAP: 0.840 -> 0.846; 0.736 -> 0.755; 0.624 -> 0.631 | Helped |
| treat low-confidence noisy labels as unlabeled rather than discarding them (ablation) | DCGMH · MIR Flickr, NUS-WIDE, and MS COCO, 40% noisy labels, 64 bits | MAP: 0.840 -> 0.846; 0.747 -> 0.755; 0.613 -> 0.631 | Mixed |
| reuse filtered noisy examples through correction and unsupervised learning rather than discarding all of them (ablation) | DCGMH · MIR Flickr, NUS-WIDE, and MS COCO, 40% noisy labels, 64 bits | MAP: 0.836 -> 0.846; 0.723 -> 0.755; 0.611 -> 0.631 | Helped |
| filter labels by distribution consistency rather than loss differences between augmented views (ablation) | DCGMH · MIR Flickr, NUS-WIDE, and MS COCO, 40% noisy labels, 64 bits | MAP: 0.828 -> 0.846; 0.701 -> 0.755; 0.622 -> 0.631 | Helped |
Surprising findings
- Performance collapses at 90% noise ratio, attributed to scarcity of clean instances for warm-up.
Limitations
- Noisy labels are synthetically injected into originally clean benchmarks rather than real annotation noise.
- Requires a warm-up phase that fails when clean instances are scarce (90% noise).
- Hyperparameters are tuned per dataset and sensitivity is only shown on MIR Flickr.
Experimental setup
Datasets: MIR Flickr NUS-WIDE MS COCO
Benchmarks: MIR Flickr NUS-WIDE MS COCO
Metrics: MAP P@N PR curve
Baselines: FOMH OMHDQ SDMH SAPMH DCMVH BSTH NCH GCIMH STBMH DIOR
Builds on: STBMH DIOR VGG
Closest papers by meaning
- Robust Self-Paced Hashing for Cross-Modal Retrieval with Noisy Labels (AAAI 2025 · similarity 0.89)
- Generalized Debiased Semi-Supervised Hashing for Large-Scale Image Retrieval (AAAI 2025 · similarity 0.87)
- Statistical Model-driven Similarity Hashing: Bridging Modalities for Efficient Unsupervised Retrieval (AAAI 2025 · similarity 0.87)
- Exploiting Descriptive Completeness Prior for Cross Modal Hashing with Incomplete Labels (NeurIPS 2024 · similarity 0.87)
- Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis (AAAI 2025 · similarity 0.87)
- Semi-Supervised Online Cross-Modal Hashing (AAAI 2025 · similarity 0.87)
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.