Research map / Architectures and efficiency / Quantization
Composite Quantization
Jingdong Wang, Ting Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| relax orthogonality to near-orthogonality constraint (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | Recall@10: 0.643 -> 0.715 | Improved |
| use near-orthogonal composite quantization instead of Cartesian k-means | Near-orthogonal composite quantization · 1M SIFT | Recall@10 (T=1, 64 bits): 63.83 -> 71.59 | Improved |
| use NOCQ instead of product quantization | Near-orthogonal composite quantization · 1M SIFT | Recall@10 (T=1, 64 bits): 60.45 -> 71.59 | Improved |
| use NOCQ instead of CKM | Near-orthogonal composite quantization · 1M CNN 64 bits | Recall@10 (T=1): 30.28 -> 36.07 | Improved |
| use NOCQ instead of CKM | Near-orthogonal composite quantization · 1B SIFT 64 bits | Recall@100 (T=1): 64.57 -> 70.12 | Improved |
| discard the third inter-dictionary term at search time (CQ-d) instead of NOCQ constraint (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | Recall@10: 0.715 -> 0.388 | Hurt |
| add near-orthogonality constraint in place of storing third term as extra 4 bytes (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | Recall@10: 0.769 -> 0.715 | Mixed |
| add near-orthogonality constraint (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | quantization error: 1.54 -> 1.65 | Hurt |
| learn epsilon instead of fixing it at zero (ablation) | Near-orthogonal composite quantization · 1M SIFT | recall | No clear effect |
| learn epsilon instead of fixing it at zero (ablation) | Near-orthogonal composite quantization · 1B SIFT | recall | Improved |
| add a global translation offset (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | recall | No clear effect |
| quantize augmented vector (QAV) instead of near-orthogonal CQ on data vectors (ablation) | Near-orthogonal composite quantization · 1M SIFT 64 bits | Recall@10: 0.715 -> 0.355 | Hurt |
Closest papers by meaning
- Norm-Explicit Quantization: Improving Vector Quantization for Maximum Inner Product Search (AAAI 2020 · similarity 0.89)
- Product Quantized Translation for Fast Nearest Neighbor Search (AAAI 2018 · similarity 0.89)
- Approximate Search with Quantized Sparse Representations (ECCV 2016 · similarity 0.89)
- Distributed Composite Quantization (AAAI 2018 · similarity 0.89)
- Polysemous Codes (ECCV 2016 · similarity 0.89)
- Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · 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.
