Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Online 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 SSOCH, a semi-supervised online cross-modal hashing method with alignment-free pseudo-labeling, Hadamard-based discriminative labels, and a tri-consistent preserving objective, evaluated on IAPR TC-12, NUS-WIDE and MIRFLICKR.
Problem
Online cross-modal hashing typically assumes fully supervised or fully unsupervised streams, ignoring streams where data mixes labeled/unlabeled and paired/unpaired multi-modal instances.
Approach
An alignment-free pseudo-labeling strategy learns per-modality feature-label projections from labeled data and transfers them to unlabeled instances without needing cross-modal pairs; Hadamard-codebook discriminative labels and fine-grained asymmetric similarity are combined in an online tri-consistency objective solved with closed-form alternating updates.
Key results
- On IAPR TC-12 Image2Text, SSOCH reaches 0.4479 mAP@all at 8 bits and 0.4992 at 64 bits, above the best baseline DOCMH (0.4234 and 0.4775).
- On MIRFLICKR Text2Image at 64 bits, SSOCH achieves 0.8037 mAP@all, the best among compared methods, vs ODCH 0.8019.
- On NUS-WIDE, SSOCH is slightly below ODCH in some settings (e.g., Text2Image 64 bits 0.7893 vs 0.7961), which the authors attribute to abundant supervision in NUS-WIDE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace one-hot labels with discriminative Hadamard embedding (ablation) | SSOCH · IAPR TC-12, Image2Text and Text2Image, 64 bits | mAP@all: 0.4463 -> 0.4992; 0.3380 -> 0.6306 | Helped |
| add pseudo-label data regularization (ablation) | SSOCH · IAPR TC-12, Image2Text and Text2Image, 64 bits | mAP@all: 0.4929 -> 0.4992; 0.6219 -> 0.6306 | Helped |
| add discriminative label embedding (ablation) | SSOCH · IAPR TC-12, Image2Text and Text2Image, 64 bits | mAP@all: 0.4942 -> 0.4992; 0.6225 -> 0.6306 | Helped |
| add asymmetric similarity preserving (ablation) | SSOCH · IAPR TC-12, Image2Text and Text2Image, 64 bits | mAP@all: 0.3283 -> 0.4992; 0.3380 -> 0.6306 | Helped |
| use SSOCH instead of DOCMH | online cross-modal hashing · NUSWIDE Image2Text, 64 bits | mAP@all: 0.6585 -> 0.6483 | Hurt |
| use SSOCH instead of ODCH | online cross-modal hashing · NUSWIDE Text2Image, 64 bits | mAP@all: 0.7961 -> 0.7893 | Hurt |
Surprising findings
- On NUS-WIDE, SSOCH is slightly inferior to the fully supervised ODCH baseline, which the authors attribute to abundant supervised semantic information in that dataset.
Limitations
- Hash codes for unlabeled data are not well learned; only supervised data's codes are stored in the retrieval library.
- Pseudo-label category count is fixed at 3 for simplicity.
- Supervised baselines are given 10% labeled instances per chunk, a simulated semi-supervised setting.
Experimental setup
Datasets: IAPR TC-12 NUS-WIDE MIRFLICKR
Benchmarks: IAPR TC-12 NUS-WIDE MIRFLICKR
Metrics: mAP@all mAP@K precision@K training time
Baselines: OCMH OCMFH OUCDH OMGH SPOCH DPOCH OLSH ROHLSE OFDDH OSAH DOCH DOCMH
Builds on: OLSH ODCH DOCMH Hadamard matrix label embedding
Closest papers by meaning
- PEOCH: Online Cross-Modal Hashing with Semi-Supervised Streaming Data Driving Prototype Evolution (AAAI 2026 · similarity 0.94)
- Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (AAAI 2025 · similarity 0.89)
- Dual Self-Paced Cross-Modal Hashing (AAAI 2024 · similarity 0.89)
- SHE: Streaming-media Hashing Retrieval (ICML 2025 · similarity 0.88)
- Robust Self-Paced Hashing for Cross-Modal Retrieval with Noisy Labels (AAAI 2025 · similarity 0.88)
- Vision-guided Text Mining for Unsupervised Cross-modal Hashing with Community Similarity Quantization (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.