Research map / Label-efficient and robust learning / Recognition and long-tail
End-To-End Supervised Product Quantization for Image Search and Retrieval
Check 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 |
|---|---|---|---|
| Train end-to-end supervised product quantization instead of using standard unsupervised PQ | DPQ · CIFAR-10 Protocol 1, 24-bit | mAP: 0.295 -> 0.7528 | Helped |
| Use DPQ rather than SUBIC | CIFAR-10 Protocol 1, 12-bit · CIFAR-10 Protocol 1, 12-bit | mAP: 0.6349 -> 0.7410 | Helped |
| Use DPQ instead of DSDH | CIFAR-10 Protocol 2 · CIFAR-10 Protocol 2, 16-bit | mAP: 0.935 -> 0.9507 | Helped |
| Use DPQ instead of HDT | ImageNet-100 · ImageNet-100, 16-bit | mAP@1000: 0.838 -> 0.886 | Helped |
| Add intra-normalization to the soft and hard sub-vectors (ablation) | DPQ, cross-domain retrieval · VOC2007 and Caltech-101 | mAP: VOC2007: 0.5371 -> 0.5647; Caltech-101: 0.4073 -> 0.4231 | Helped |
| Add intra-normalization to the soft and hard sub-vectors (ablation) | DPQ, cross-domain retrieval · ImageNet | mAP: 0.3231 -> 0.3227 | No clear effect |
| Apply L2 normalization during model training instead of post-hoc intra-normalization (ablation) | DPQ, cross-domain retrieval · Caltech-101 and VOC2007 | mAP | Hurt |
| Increase the joint central loss weight from zero to 0.1 (ablation) | DPQ, cross-domain retrieval · VOC2007, Caltech-101, and ImageNet | mAP | Helped |
| Increase the joint central loss weight beyond 0.1 (ablation) | DPQ, cross-domain retrieval · VOC2007, Caltech-101, and ImageNet | mAP | Hurt |
| Use asymmetric rather than symmetric search (ablation) | DPQ, cross-domain retrieval · VOC2007, Caltech-101, and ImageNet | mAP: VOC2007: 0.5340 -> 0.5371; Caltech-101: 0.4035 -> 0.4073; ImageNet: 0.3183 -> 0.3231 | Helped |
| Use the 3-layer representation instead of the 2-layer representation (ablation) | DPQ, cross-domain retrieval · ImageNet | mAP: 0.3231 -> 0.3532 | Helped |
| Use PQ-Norm rather than standard PQ | Product Quantization · Oxford5K | mAP: 0.2374 -> 0.2646 $\pm$ 0.0012 | Helped |
Closest papers by meaning
- Unleashing the Full Potential of Product Quantization for Large-Scale Image Retrieval (NeurIPS 2023 · similarity 0.93)
- Deep Quantization Network for Efficient Image Retrieval (AAAI 2016 · similarity 0.89)
- One Loss for Quantization: Deep Hashing With Discrete Wasserstein Distributional Matching (CVPR 2022 · similarity 0.88)
- HiHPQ: Hierarchical Hyperbolic Product Quantization for Unsupervised Image Retrieval (AAAI 2024 · similarity 0.87)
- Pairwise-Label-Based Deep Incremental Hashing with Simultaneous Code Expansion (AAAI 2024 · similarity 0.85)
- Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization (ICML 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.