Research map / Vision-language and multimodal / Cross-modal learning
Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval
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 LCDH, a knowledge-distillation framework for online cross-modal hashing where a CLIP-plus-attention offline teacher supervises a lightweight VGG16+BoW online student via similarity matrix approximation.
Problem
Online cross-modal hashing for real-time data streams struggles to capture coexistent cross-modal semantic relevance and to transfer knowledge from strong offline models to lightweight online student networks.
Approach
Bridges offline and online cross-modal hashing by distilling a CLIP-based teacher's similarity matrix (with attention-based feature enhancement) into a lightweight VGG16/BoW student through similarity matrix approximation with a log-likelihood similarity preserving term.
Key results
- On MIRFlickr-25K at 128 bits, LCDH reaches 0.8041 mAP for I→T and 0.8574 for T→I, above the best baseline ROH (0.7626 and 0.8481).
- On NUS-WIDE at 128 bits, LCDH reaches 0.8015 (I→T) and 0.8311 (T→I) mAP, versus best baselines of 0.7001 (ROH) and 0.8264 (ROH).
- Ablation shows removing the CLIP features or the offline teacher each reduces mAP, with the offline-hashing removal causing a drop of about 0.03 on MIRFlickr I→T at 128 bits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add log-likelihood similarity preservation (ablation) | LCDH student network · MIRFlickr-25K, IAPR TC-12, and NUS-WIDE; I→T and T→I, 128 bits | mAP: 0.7981 -> 0.8041; 0.6253 -> 0.6499; 0.7991 -> 0.8015; 0.8532 -> 0.8574; 0.6598 -> 0.6810; 0.8212 -> 0.8311 | Helped |
| Add offline teacher hashing (ablation) | LCDH · MIRFlickr-25K, IAPR TC-12, and NUS-WIDE; I→T and T→I, 128 bits | mAP: 0.7702 -> 0.8041; 0.6004 -> 0.6499; 0.7865 -> 0.8015; 0.8395 -> 0.8574; 0.6444 -> 0.6810; 0.8090 -> 0.8311 | Helped |
| Add CLIP-based feature extraction (ablation) | LCDH teacher network · MIRFlickr-25K, IAPR TC-12, and NUS-WIDE; I→T and T→I, 128 bits | mAP: 0.7716 -> 0.8041; 0.6107 -> 0.6499; 0.7906 -> 0.8015; 0.8469 -> 0.8574; 0.6585 -> 0.6810; 0.8133 -> 0.8311 | Helped |
Surprising findings
- Baseline results on IAPR TC-12 are much lower than on the other datasets for all methods, which the authors attribute to more categories and feature dimensions.
Limitations
- Experiments use only three datasets and no timing or memory measurements to substantiate the lightweight claim.
- Hyperparameter sensitivity is shown only on MIRFlickr-25K, with a large search grid (1e-5 to 1e5).
Experimental setup
Datasets: MIRFlickr-25K IAPR TC-12 NUS-WIDE
Benchmarks: MIRFlickr-25K IAPR TC-12 NUS-WIDE
Metrics: mAP precision-recall curve top-N precision
Baselines: DCMH SSAH UKD DCHUC DAEH SKDCH OLSH OCMFH LEMON SDOH ROH
Builds on: CLIP UKD SKDCH LEMON ROH
Closest papers by meaning
- An End-To-End Graph Attention Network Hashing for Cross-Modal Retrieval (NeurIPS 2024 · similarity 0.92)
- Dynamic Masking and Auxiliary Hash Learning for Enhanced Cross-Modal Retrieval (NeurIPS 2025 · similarity 0.92)
- Stationary and Clustering Transformer Hashing for Cross-modal Retrieval (AAAI 2026 · similarity 0.91)
- Vision-guided Text Mining for Unsupervised Cross-modal Hashing with Community Similarity Quantization (AAAI 2025 · similarity 0.91)
- Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis (AAAI 2025 · similarity 0.90)
- Statistical Model-driven Similarity Hashing: Bridging Modalities for Efficient Unsupervised Retrieval (AAAI 2025 · similarity 0.90)
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.