Research map / Architectures and efficiency / Quantization
Differentiable Product Quantization for Memory Efficient Camera Relocalization
Zakaria Laskar, Iaroslav Melekhov, Assia Benbihi, Shuzhe Wang, Juho Kannala
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 |
|---|---|---|---|
| add decoder trained with triplet loss instead of L2 loss (ablation) | PQ4 + decoder · Aachen Day-Night night | accuracy % (0.25m,2deg): 70.4 -> 79.6 | Improved |
| add differentiable PQ encoder (learned codebooks) (ablation) | PQ4 · Aachen Day-Night night | accuracy % (0.25m,2deg): 45.9 -> 68.4 | Improved |
| add D-PQED dequantization to PQ plus map compression at 1MB | PQ4+Gamma · Aachen Day-Night night | accuracy % (0.25m,2deg): 36.7 -> 77.6 | Improved |
| also quantize and dequantize query descriptors (symmetric setup) (ablation) | D-PQED · Aachen Day-Night night | accuracy % (0.25m,2deg): 81.6 -> 74.5 | Hurt |
| use N-pair loss instead of triplet loss (ablation) | D-PQED · Aachen Day-Night night | accuracy % (0.25m,2deg): 79.6 -> 76.5 | Hurt |
| use mutual nearest-neighbor matcher instead of SuperGlue (ablation) | D-PQED · Aachen Day-Night night | accuracy % (0.25m,2deg): 81.6 -> 38.8 | Hurt |
| balance PQ level M and map compression alpha at fixed memory (PQ4 vs PQ32 or PQ2) (ablation) | PQ+Gamma+D-PQED · Aachen Day-Night night, 1MB | accuracy % (0.25m,2deg): PQ32 62.2, PQ4 77.6, PQ2 66.3 | Mixed |
| descriptor product quantization alone (PQ32) versus map compression at the same memory (ablation) | PQ vs Gamma · Aachen Day-Night day | accuracy % (0.25m,2deg): 83.9 -> 88.6 | Improved |
Closest papers by meaning
- Simple but Effective Triplet-Based Compression Strategies for Compact Visual Localization (CVPR 2026 · similarity 0.88)
- MAD-DR: Map Compression for Visual Localization with Matchness Aware Descriptor Dimension Reduction (ECCV 2024 · similarity 0.87)
- SceneSqueezer: Learning To Compress Scene for Camera Relocalization (CVPR 2022 · similarity 0.87)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.85)
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression (CVPR 2025 · similarity 0.83)
- NeRFCodec: Neural Feature Compression Meets Neural Radiance Fields for Memory-Efficient Scene Representation (CVPR 2024 · similarity 0.83)
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.
