Research map / Label-efficient and robust learning / Semi-supervised learning
SSAH: Semi-Supervised Adversarial Deep Hashing with Self-Paced Hard Sample Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSAH instead of SSGAH | Semi-supervised hashing · CIFAR-10 and NUS-WIDE, 12/24/48 bits | mAP: 0.819 -> 0.862; 0.837 -> 0.878; 0.855 -> 0.886; 0.835 -> 0.872; 0.847 -> 0.884; 0.865 -> 0.898 | Helped |
| Use SSAH instead of DSaH | Fine-grained image hashing · Stanford Dogs-120 and CUB Bird, 12/24/48 bits | mAP: 0.244 -> 0.273; 0.287 -> 0.343; 0.408 -> 0.478; 0.091 -> 0.141; 0.2087 -> 0.265; 0.285 -> 0.359 | Helped |
| Use SSAH instead of SSGAH | Hashing for unseen-class retrieval · CIFAR-10 and NUS-WIDE unseen classes, 12/24/48 bits | mAP: 0.309 -> 0.338; 0.323 -> 0.370; 0.339 -> 0.379; 0.539 -> 0.569; 0.553 -> 0.571; 0.579 -> 0.596 | Helped |
| Add MARN to the baseline (ablation) | AlexNet hashing · CIFAR-10 and NUS-WIDE, 12/48 bits | mAP: 0.751 -> 0.801; 0.792 -> 0.831; 0.762 -> 0.810; 0.801 -> 0.842 | Helped |
| Add MSMN to the baseline (ablation) | AlexNet hashing · CIFAR-10 and NUS-WIDE, 12/48 bits | mAP: 0.751 -> 0.830; 0.792 -> 0.853; 0.762 -> 0.840; 0.801 -> 0.861 | Helped |
| Combine MARN and MSMN rather than random rotation and masking (ablation) | AlexNet hashing · CIFAR-10 and NUS-WIDE, 12/48 bits | mAP: 0.796 -> 0.862; 0.816 -> 0.886; 0.810 -> 0.872; 0.833 -> 0.898 | Helped |
| Add the semi-supervised consistent loss to H-Net (ablation) | SSAH with A-Net trained using l_sem+l_sp · CIFAR-10, 12/24/48 bits | mAP: 0.841 -> 0.862; 0.855 -> 0.878; 0.867 -> 0.886 | Helped |
| Add self-paced adversarial loss to H-Net (ablation) | SSAH H-Net · CIFAR-10, 12/24/48 bits | mAP: 0.862 -> 0.843; 0.878 -> 0.852; 0.886 -> 0.8612 | Hurt |
| Replace self-paced loss with fixed-paced loss (ablation) | SSAH · CIFAR-10, 12/48 bits, ω=0.1 | mAP: 0.862 -> 0.839; 0.886 -> 0.861 | Hurt |
| Use self-paced rather than fixed-paced loss across reported margin settings (ablation) | SSAH · CIFAR-10, 12/48 bits | mAP: 0.790 vs 0.791; 0.803 vs 0.815; 0.843 vs 0.819; 0.855 vs 0.842; 0.855 vs 0.838; 0.873 vs 0.852; 0.846 vs 0.789; 0.861 vs | Mixed |
Closest papers by meaning
- Unsupervised Deep Generative Adversarial Hashing Network (CVPR 2018 · similarity 0.89)
- AutoSSVH: Exploring Automated Frame Sampling for Efficient Self-Supervised Video Hashing (CVPR 2025 · similarity 0.87)
- Deep Graph Online Hashing for Multi-Label Image Retrieval (AAAI 2025 · similarity 0.84)
- Efficient Self-Supervised Video Hashing with Selective State Spaces (AAAI 2025 · similarity 0.84)
- Generalized Debiased Semi-Supervised Hashing for Large-Scale Image Retrieval (AAAI 2025 · similarity 0.84)
- HUANG: A Robust Diffusion Model-based Targeted Adversarial Attack Against Deep Hashing Retrieval (AAAI 2025 · similarity 0.84)
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.