Research map / Optimization / Optimizers
SDGMNet: Statistic-Based Dynamic Gradient Modulation for Local Descriptor Learning
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 SDGMNet, a dynamic gradient-modulation loss for contrastive local descriptor learning that uses running statistics of angular distances (auto-focus modulation, probabilistic margin, power adjustment) to improve patch verification, matching, retrieval, pose estimation and 3D reconstruction.
Problem
Static gradient modulation in contrastive local descriptor learning ignores global gradient statistics that change across training phases and datasets, making learned descriptors sensitive to these shifts.
Approach
Uses included angle as the distance measure (its backpropagated gradient magnitude depends only on the norm), then modulates pair gradients with Gaussian-blur self weights, a CDF-based probabilistic hard margin, and a power-normalization step, all driven by dynamically estimated (Bayesian running) statistics.
Key results
- On UBC PhotoTour, SDGMNet achieves mean FPR@95 of 0.78% vs. HyNet's 0.84%.
- On IMC PhotoTourism, SDGMNet raises mean mAA from 64.73 (HyNet) to 65.70, and DISK+SDGM improves from 72.02 to 72.65 mean mAA.
- On ScanNet, DISK+SDGM improves AUC@20 from 27.23 to 32.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDGMNet instead of HyNet | local descriptor learning · UBC PhotoTour | FPR@95 (%): 0.84 -> 0.78 | Helped |
| Use SDGMNet instead of HardNet | local descriptors · IMC PhotoTourism validation | mean mAA (%): 64.32 -> 65.70 | Helped |
| Fine-tune DISK with SDGM loss | DISK · IMC PhotoTourism validation | mean mAA (%): 72.02 -> 72.65 | Helped |
| Fine-tune ALIKE with SDGM loss | ALIKE · IMC PhotoTourism validation | mean mAA (%): 65.12 -> 66.33 | Helped |
| Fine-tune DISK with SDGM loss | DISK · ScanNet | AUC at 20°: 27.23 -> 32.95 | Helped |
| Use SDGMNet instead of HardNet | local descriptors · ScanNet | AUC at 20°: 39.72 -> 41.51 | Helped |
| Use SDGMNet instead of HyNet | local descriptors · ETH 3D reconstruction, Madrid Metropolis | registered images: 807 -> 850 | Helped |
| Use SDGMNet instead of HyNet | local descriptors · ETH 3D reconstruction, Gendarmenmarkt | registered images: 1171 -> 1184 | Helped |
| Add probabilistic hard margin (ablation) | SDGMNet · UBC PhotoTour | FPR@95 (%): about 0.18 | Helped |
| Add power adjustment (ablation) | SDGMNet modulation frameworks · UBC PhotoTour | FPR@95 (%) | Helped |
| Use angular distance rather than cosine or L2 distance (ablation) | SDGMNet modulation · UBC PhotoTour | FPR@95 (%) | Helped |
| Add auto-focus modulation (ablation) | SDGMNet · UBC PhotoTour | FPR@95 (%) | Helped |
Surprising findings
- Angular distance and auto-focus modulation alone give little benefit without the power adjustment step, which the authors attribute to bias introduced by dropping scale factors.
Limitations
- Descriptor evaluation is limited to 128-d patch descriptors trained on UBC PhotoTour with a fixed HyNet-style architecture.
- Several hyperparameters (m, alpha, initial powers) are set empirically; sensitivity shown only on Liberty.
- Gains over HyNet are often small (e.g., about 0.06 FPR@95 on mean, and no gap on some matching subtasks).
Experimental setup
Datasets: UBC PhotoTour HPatches Image Matching Challenge (PhotoTourism) ScanNet ETH3D MegaDepth
Benchmarks: UBC PhotoTour HPatches Image Matching Challenge ScanNet ETH3D
Metrics: FPR@95 mAP mAA@10 AUC registered images sparse points dense points
Baselines: SIFT L2Net HardNet CDFDesc SOSNet HyNet PUDesc DISK ALIKE AWDesc
Builds on: HardNet HyNet CDFDesc Multi-Simi Loss Circle Loss
Closest papers by meaning
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.84)
- A Stochastic Approach to Bi-Level Optimization for Hyperparameter Optimization and Meta Learning (AAAI 2025 · similarity 0.84)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.83)
- Towards Interpretable Deep Local Learning with Successive Gradient Reconciliation (ICML 2024 · similarity 0.83)
- DenoiseRep: Denoising Model for Representation Learning (NeurIPS 2024 · similarity 0.82)
- Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs (ICLR 2024 · similarity 0.82)
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.