Research map / Graph learning / GNN architectures
Deep Graph Online Hashing for Multi-Label Image 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 Deep Graph Online Hashing (DGOH), which uses inductive GNNs over feature and label graphs with learned label-wise weights to build the affinity matrix, enabling deep online hashing for multi-label image retrieval.
Problem
Online hashing for streaming image retrieval relies on shallow models because deep models are too slow to update, and existing methods under-use multi-label supervision and keep the affinity matrix fixed.
Approach
Introduces inductive GNNs into online hashing: a feature graph (image features, edges from the affinity matrix) and a label graph (word2vec label embeddings, edges from label co-occurrence) whose learned label-wise weights dynamically update the affinity matrix; an old-sample selection strategy mitigates catastrophic forgetting.
Key results
- DGOH achieves up to 11.6% mAP gain on MIRFlickr and 13.3% on MS-COCO over the best baselines across 16-128 bits.
- DGOH at 32 bits reaches 0.875 mAP on MIRFlickr versus 0.768 for FOH, with 0.659 on MS-COCO versus 0.531 for FOH.
- Training time for DGOH at 32 bits is about 391 s on MIRFlickr and 877 s on MS-COCO, higher than baselines but described as competitive.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the DGOH method instead of FOH | online hashing · MIRFlickr | mAP (32 bits): 0.768 -> 0.875 | Helped |
| use the DGOH method instead of FOH | online hashing · MS-COCO | mAP (32 bits): 0.531 -> 0.659 | Helped |
| add the label network (ablation) | DGOH without the label network or label-wise weights · MIRFlickr | mAP (32 bits): 0.859 -> 0.871 | Helped |
| add the label network (ablation) | DGOH without the label network or label-wise weights · MS-COCO | mAP (32 bits): 0.618 -> 0.621 | Helped |
| use learned label-wise weights instead of initialized weights (ablation) | DGOH · MIRFlickr | mAP (32 bits): 0.871 -> 0.875 | Helped |
| use learned label-wise weights instead of initialized weights (ablation) | DGOH · MS-COCO | mAP (32 bits): 0.640 -> 0.659 | Helped |
| replace label-wise affinity with binary shared-label affinity (ablation) | DGOH · MIRFlickr | mAP (32 bits): 0.875 -> 0.866 | Hurt |
| replace label-wise affinity with binary shared-label affinity (ablation) | DGOH · MS-COCO | mAP (32 bits): 0.659 -> 0.627 | Hurt |
| construct the feature graph with a fixed number of neighbors (ablation) | DGOH · MIRFlickr | mAP (32 bits): 0.875 -> 0.826 | Hurt |
| construct the feature graph with a fixed number of neighbors (ablation) | DGOH · MS-COCO | mAP (32 bits): 0.659 -> 0.673 | Helped |
| remove thresholding by λ in feature graph construction (ablation) | DGOH · MIRFlickr | mAP (32 bits): 0.875 -> 0.869 | Hurt |
| remove thresholding by λ in feature graph construction (ablation) | DGOH · MS-COCO | mAP (32 bits): 0.659 -> 0.634 | Hurt |
Surprising findings
- The variant without the additional losses still outperforms all shallow baselines, suggesting the deep model alone drives much of the gain.
- Deep graph online hashing training is substantially slower than shallow baselines (e.g., 391 s vs 6.87 s for OKH on MIRFlickr), a trade-off the authors accept.
Limitations
- Training time is higher than most shallow online hashing baselines.
- Batch size must be fixed at 50 for the inductive GNN graph construction.
- Hyperparameter sensitivity is only tested on MIRFlickr at 32 bits.
Experimental setup
Datasets: MIRFlickr MS-COCO
Benchmarks: MIRFlickr MS-COCO
Metrics: mAP Precision@500 training time query time
Baselines: OKH AdaptHash OSupH MIHash BSODH HMOH SDOH FOH
Builds on: GraphSAGE word2vec ViG (vision GNN) ODHUC FOH
Closest papers by meaning
- Deep Incremental Hashing Network for Efficient Image Retrieval (CVPR 2019 · similarity 0.90)
- Deep Hashing Network for Efficient Similarity Retrieval (AAAI 2016 · similarity 0.89)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.89)
- Deep Supervised Discrete Hashing (NeurIPS 2017 · similarity 0.89)
- Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.87)
- Generalized Debiased Semi-Supervised Hashing for Large-Scale Image Retrieval (AAAI 2025 · similarity 0.86)
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.