Research map / Label-efficient and robust learning / Recognition and long-tail
Characteristics Matching Based Hash Codes Generation for Efficient Fine-grained 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 CMBH, a hashing method for efficient fine-grained image retrieval that generates binary codes by matching features against learned characteristic vectors (C-vectors), with cross-layer transfer and multi-region training modules used only during training.
Problem
Efficient fine-grained image retrieval with short binary hash codes struggles because low-dimensional codes cannot embed the detailed features needed for fine-grained categories, and complex multi-region/multi-layer feature extractors raise inference cost.
Approach
Instead of mapping concatenated feature vectors directly to hash codes, hash bits are generated from matching scores between each sample's features and learnable per-category characteristic vectors; cross-layer semantic transfer and parameter-free multi-region embedding shape these vectors during training only, so inference stays cheap.
Key results
- On CUB-200-2011 at 12 bits, CMBH reaches 84.07 MAP versus 77.10 for the next best (CNET) with ResNet-50.
- On NABirds at 12 bits, CMBH reaches 74.42 MAP versus 68.42 for CNET.
- CMBH uses 14.3M parameters and 4.35 GFLOPs at inference on CUB, fewer than SEMICON (42.4M) and CNET (41.1M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CMBH rather than the strongest listed baseline at each code length | CUB-200-2011 retrieval · CUB-200-2011 | MAP (%): 77.10 -> 84.07; 82.11 -> 85.79; 83.09 -> 86.21; 83.92 -> 86.47 | Helped |
| Use CMBH rather than the strongest listed baseline at each code length | Aircrafts retrieval · Aircrafts | MAP (%): 88.29 -> 89.11; 89.20 -> 91.45; 89.28 -> 91.60; 89.49 -> 92.88 | Helped |
| Use CMBH rather than the strongest listed baseline at each code length | Food101 retrieval · Food101 | MAP (%): 83.06 -> 87.85; 85.85 -> 88.71; 86.35 -> 89.28; 86.42 -> 88.87 | Helped |
| Use CMBH rather than the strongest listed baseline at each code length | NABirds retrieval · NABirds | MAP (%): 68.42 -> 74.42; 75.73 -> 81.04; 77.11 -> 81.64; 78.81 -> 82.07 | Helped |
| Use CMBH rather than the strongest listed baseline at each code length | VegFru retrieval · VegFru | MAP (%): 81.63 -> 84.37; 86.41 -> 88.63; 86.80 -> 88.46; 87.75 -> 89.06 | Helped |
| Add multi-layer semantic information integration for C-vector optimization (ablation) | CMBH · CUB-200-2011 | MAP (%): 17.89 -> 77.56; 48.07 -> 81.12; 55.25 -> 81.61; 51.90 -> 82.26 | Helped |
| Add cross-layer information transfer supervision (ablation) | CMBH with multi-layer semantic integration · CUB-200-2011 | MAP (%): 77.56 -> 80.30; 81.12 -> 82.16; 81.61 -> 82.68; 82.26 -> 82.80 | Helped |
| Add multi-region feature embedding (ablation) | CMBH with multi-layer integration and cross-layer transfer · CUB-200-2011 | MAP (%): 80.30 -> 84.07; 82.16 -> 85.79; 82.68 -> 86.21; 82.80 -> 86.47 | Helped |
| Use only the basic hash-code generation and training procedure, without C-vector optimization (ablation) | CMBH · CUB-200-2011 | MAP (%): 84.07 -> 17.89; 85.79 -> 48.07; 86.21 -> 55.25; 86.47 -> 51.90 | Hurt |
| Use one C-vector per category (k=1) instead of multiple C-vectors (ablation) | CMBH · CUB-200-2011 | MAP (%) | Hurt |
| Increase the number of C-vectors per category excessively (ablation) | CMBH · CUB-200-2011 | MAP (%) | Hurt |
| Vary the gradient scale α (ablation) | CMBH · CUB-200-2011 | MAP (%) | Mixed |
Surprising findings
- Simply adding more high-dimensional fine-grained feature vectors does not reliably improve performance once mapped to short binary codes, due to information loss in the mapping.
Limitations
- Adaptive scaling parameter alpha must be tuned per code length and is left for future work.
- Multi-region cropping uses a fixed M=4 chosen empirically.
Experimental setup
Datasets: CUB-200-2011 FGVC-Aircrafts Food101 NABirds VegFru
Benchmarks: CUB-200-2011 FGVC-Aircrafts Food101 NABirds VegFru
Metrics: MAP@all parameters (M) FLOPs (G)
Baselines: ExchNet A2-Net A2-Net++ FCAENet SEMICON FISH AGMH CNET DLTH sRLH
Builds on: SCDA knowledge distillation NMS
Closest papers by meaning
- Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization (ICML 2025 · similarity 0.90)
- An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing (CVPR 2024 · similarity 0.86)
- Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.85)
- Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis (AAAI 2025 · similarity 0.85)
- Statistical Model-driven Similarity Hashing: Bridging Modalities for Efficient Unsupervised Retrieval (AAAI 2025 · similarity 0.85)
- Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.85)
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.