Research map / Continual learning / Class-incremental
Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion
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 CEDIH, a triple-channel asymmetric deep hashing framework that learns new classes while expanding hash code length, expanding existing database codes via a learned projection matrix instead of regenerating them with the CNN.
Problem
Deep incremental hashing methods can learn new classes without changing old codes, but when more classes require longer codes, the entire database must be re-encoded; existing methods cannot expand code length and learn new classes simultaneously.
Approach
A triple-channel asymmetric framework trains a new CNN with a target code length plus a code projection matrix that maps old database codes to new bits, with a pairwise-label incremental similarity-preserving loss and a double-end quantization loss to preserve new-new, original-original and original-new similarities.
Key results
- CEDIH+ADSH reaches 0.7898 mAP on CIFAR-10@ALL at 48 bits versus 0.7787 for ADSH, and 0.8481 vs 0.8396 on NUS-WIDE@5000 at 48 bits.
- Code expansion time on ImageNet is about 46.5 s for CEDIH+ADSH versus about 153.7 s for ADSH at 48 bits, roughly four times faster.
- On NUS-WIDE, CEDIH+ADSH at 24 bits is 0.8328 vs DIHN+ADSH 0.8228 under setting 3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CEDIH to ADSH | ADSH · CIFAR-10@ALL, 48 bits | mAP: 0.7787 -> 0.7898 | Helped |
| apply CEDIH to ADSH | ADSH · NUS-WIDE@5000, 48 bits | mAP: 0.8396 -> 0.8481 | Helped |
| apply CEDIH to ADSH | ADSH · ImageNet@1000, 48 bits | mAP: 0.6653 -> 0.6728 | Helped |
| add the double-end quantization loss (ablation) | CEDIH with ADSH · CIFAR-10, setting 3, 48 bits | mAP: 0.7804 -> 0.7898 | Helped |
| add the double-end quantization loss (ablation) | CEDIH with ADSH · NUS-WIDE, setting 3, 48 bits | mAP: 0.8398 -> 0.8481 | Helped |
| add the double-end quantization loss (ablation) | CEDIH with ADSH · ImageNet, setting 3, 48 bits | mAP: 0.6659 -> 0.6728 | Helped |
| retain all reported quantization terms rather than remove one (ablation) | CEDIH with ADSH · CIFAR-10, setting 3, 48 bits | mAP: 0.7836 -> 0.7898; 0.7830 -> 0.7898; 0.7823 -> 0.7898 | Helped |
| retain all reported quantization terms rather than remove one (ablation) | CEDIH with ADSH · NUS-WIDE, setting 3, 48 bits | mAP: 0.8398 -> 0.8481; 0.8420 -> 0.8481; 0.8403 -> 0.8481 | Mixed |
| retain all reported quantization terms rather than remove one (ablation) | CEDIH with ADSH · ImageNet, setting 3, 48 bits | mAP: 0.6716 -> 0.6728; 0.6705 -> 0.6728; 0.6690 -> 0.6728 | Helped |
| use incremental code expansion with CEDIH | ADSH · ImageNet, setting 3, code expansion | code expansion time (seconds): 153.70 -> 46.50 | Helped |
Surprising findings
- On ImageNet under settings 1 and 2, CEDIH underperforms its ADSH backbone, which the authors attribute to limited semantic capacity of new bits generated from fixed original codes.
Limitations
- New bits generated from fixed original database codes have limited semantic capacity, causing underperformance vs backbone on ImageNet settings 1 and 2.
- Evaluated only on image retrieval with pairwise-label backbones; comparisons limited to CIFAR-10, NUS-WIDE and ImageNet.
Experimental setup
Datasets: CIFAR-10 NUS-WIDE ImageNet
Benchmarks: CIFAR-10 NUS-WIDE ImageNet
Metrics: mAP precision-recall curves code expansion time
Baselines: SH ITQ LFH SDH DPSH HashNet DSDH ADSH DIHN
Builds on: ADSH DIHN CEDH DPSH
Closest papers by meaning
- Deep Incremental Hashing Network for Efficient Image Retrieval (CVPR 2019 · similarity 0.92)
- One Loss for All: Deep Hashing with a Single Cosine Similarity based Learning Objective (NeurIPS 2021 · similarity 0.91)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.91)
- Deep Hashing Network for Efficient Similarity Retrieval (AAAI 2016 · similarity 0.90)
- Deep Graph Online Hashing for Multi-Label Image Retrieval (AAAI 2025 · similarity 0.87)
- Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization (ICML 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.