Research map / Label-efficient and robust learning / Semi-supervised learning
Generalized Debiased Semi-Supervised Hashing for Large-Scale 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 GDSH, a pseudo-label-based semi-supervised hashing method that reduces bias via robust label completion, debiased hash learning (label denoising and similarity updating), and discrete cyclic optimization.
Problem
Semi-supervised hashing with pseudo labels suffers biased learning from pseudo-label noise, hash learning objectives, and relaxation-induced quantization error during optimization, limiting retrieval accuracy on large-scale image data.
Approach
Decouples semi-supervised hashing into robust label completion (l2,1-regularized linear label projection), debiased hash learning that sparsely denoises pseudo labels via an l1 error term and updates pairwise similarities at the sample level, and discrete cyclic coordinate descent for binary codes, avoiding costly sample weighting.
Key results
- On ImageNet with 30% supervision, GDSH surpasses the best semi-supervised baseline by over 1.4%, 0.4%, 4.5%, and 6.6% mAP at 16, 32, 48, and 64 bits.
- On NUS-WIDE, GDSH improves mAP by over 3.3% with 48-bit codes over the best baseline.
- On MIRFlickr with CLIP features, GDSH reaches 0.9359 mAP at 48 bits versus 0.8996 for MLAGH.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing · CALTECH-101, 30% supervision, 64 bits | mAP: 0.4956 -> 0.5163 | Helped |
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing · CIFAR-10, 30% supervision, 64 bits | mAP: 0.5994 -> 0.6565 | Helped |
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing · ImageNet, 30% supervision, 64 bits | mAP: 0.3348 -> 0.3473 | Helped |
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing · MS-COCO, 30% supervision, 48 bits | mAP: 0.8584 -> 0.8678 | Helped |
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing · NUS-WIDE, 30% supervision, 48 bits | mAP: 0.8038 -> 0.8047 | Helped |
| Use GDSH instead of the strongest listed baseline | semi-supervised hashing with CLIP features · MIRFlickr, 30% supervision, 48 bits | mAP: 0.8996 -> 0.9359 | Helped |
| Add the second and third terms in the robust label completion objective (ablation) | RLC · CIFAR-10, 30% supervision, 64 bits | mAP: 63.90 -> 65.90 | Helped |
| Add the third term in the robust label completion objective (ablation) | RLC · CIFAR-10, 30% supervision, 64 bits | mAP: 65.39 -> 65.90 | Helped |
| Add the second term in the robust label completion objective (ablation) | RLC · CIFAR-10, 30% supervision, 64 bits | mAP: 65.86 -> 65.90 | Helped |
| Add debiasing to hash learning (ablation) | DHL · CIFAR-10, 30% supervision, 64 bits | mAP: 64.00 -> 65.90 | Helped |
| Add similarity updating (ablation) | DHL · CIFAR-10, 30% supervision, 64 bits | mAP: 64.52 -> 65.90 | Helped |
| Add label denoising (ablation) | DHL · CIFAR-10, 30% supervision, 64 bits | mAP: 65.24 -> 65.90 | Helped |
Surprising findings
- Assigning only a single pseudo label (top-1) already yields competitive performance on NUS-WIDE, as shown in the parameter sensitivity analysis.
Limitations
- Number of pseudo labels k is fixed by hand (k=1 for single-label, k=2 for multi-label) rather than adaptively determined.
- Relies on a linear hash function for out-of-sample extension and linear label projection, which may limit expressiveness.
- Sample-level debiasing uses a random subset (beta=0.3) of similarities for tractability.
Experimental setup
Datasets: CALTECH-101 CIFAR-10 ImageNet MS-COCO NUS-WIDE MIRFlickr
Benchmarks: CALTECH-101 CIFAR-10 ImageNet MS-COCO NUS-WIDE MIRFlickr
Metrics: mAP Precision
Baselines: SH PCA-ITQ PCA-RR MFH SDH NSH FSDH SDHMLR SCDH LBSE DSPH REPH
Builds on: Supervised Discrete Hashing (SDH) Semi-supervised hashing (SSH) Natural Supervised Hashing (NSH) Fast Supervised Discrete Hashing (FSDH) l2,1-norm robust feature selection
Closest papers by meaning
- Supervised Deep Hashing for Hierarchical Labeled Data (AAAI 2018 · similarity 0.88)
- Discrete Image Hashing Using Large Weakly Annotated Photo Collections (AAAI 2016 · similarity 0.88)
- Deep Supervised Discrete Hashing (NeurIPS 2017 · similarity 0.88)
- Distribution-Consistency-Guided Multi-modal Hashing (AAAI 2025 · similarity 0.87)
- Deep Graph Online Hashing for Multi-Label Image Retrieval (AAAI 2025 · similarity 0.86)
- Semi-Supervised Online Cross-Modal Hashing (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.