Research map / Architectures and efficiency / CNN design
Metric Convolutions: A Unifying Theory to Adaptive Image Convolutions
Thomas Dagès, Michael Lindenbaum, Alfred M. Bruckstein
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 |
|---|---|---|---|
| replace deformable convolution with metric UTB convolution | single-layer denoiser · single image denoising (sigma 0.3, k=5) | normalised generalisation gap: 265 -> 1.3 | Improved |
| replace deformable convolution with metric UTB convolution at large kernel | single-layer denoiser · single image denoising (sigma 0.3, k=121) | test MSE: 6.18e-2 -> 7.82e-3 | Improved |
| increase metric asymmetry (eps_w 0.9 to 0.1) (ablation) | metric UTB convolution · single image denoising (sigma 0.3, k=11) | test MSE: 7.43e-3 -> 7.32e-3 | Improved |
| replace deformable with metric UTB convolution (fixed weights, scratch) | ResNet18 · CIFAR-10 | top-1 accuracy: 33.21% -> 69.83% | Improved |
| replace deformable with metric UTB convolution (fixed weights, scratch) | ResNet18 · CIFAR-100 | top-1 accuracy: 10.47% -> 40.63% | Improved |
| replace deformable with metric UTB convolution (learnable weights, scratch) | ResNet18 · CIFAR-100 | top-1 accuracy: 54.51% -> 70.03% | Improved |
| replace standard convolution with metric UTB (learnable weights, transfer) | ResNet18 · CIFAR-10 | top-1 accuracy: 92.64% -> 93.07% | Improved |
| replace standard convolution with metric UTB (learnable weights, transfer) | ResNet18 · CIFAR-100 | top-1 accuracy: 70.52% -> 70.38% | No clear effect |
| replace standard convolution with metric UTB (learnable weights, transfer) | ResNet18 · Fashion-MNIST | top-1 accuracy: 93.30% -> 93.49% | Improved |
| replace deformable with metric UTB on CIFAR-10 LKW-TL | ResNet18 · CIFAR-10 | top-1 accuracy: 93,10% -> 93.07% | No clear effect |
| learn kernel weights instead of fixing them (from scratch) (ablation) | metric UTB ResNet18 · CIFAR-10 | top-1 accuracy: 69.83% -> 91.83% | Improved |
| use transfer learning instead of scratch (learnable weights) (ablation) | metric UTB ResNet18 · CIFAR-100 | top-1 accuracy: 70.03% -> 70.38% | No clear effect |
Closest papers by meaning
- Learning Low-Dimensional Metrics (NeurIPS 2017 · similarity 0.87)
- Signal-To-Noise Ratio: A Robust Distance Metric for Deep Metric Learning (CVPR 2019 · similarity 0.87)
- Hyperbolic Vision Transformers: Combining Improvements in Metric Learning (CVPR 2022 · similarity 0.86)
- The Fault in Our Metrics: Revisiting Generative Model Evaluation with Chamfer Distance (NeurIPS 2026 · similarity 0.86)
- Deep Metric Learning with Adaptively Composite Dynamic Constraints (TPAMI 2023 · similarity 0.86)
- Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces (NeurIPS 2026 · similarity 0.86)
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.
