Research map / Image generation / Image restoration
Neural Nearest Neighbors Networks
Tobias Plötz, Stefan Roth
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 end-to-end learned continuous nearest-neighbor processing | DnCNN image denoising · Urban100, σ=25 | PSNR: 29.97 -> 30.19 | Improved |
| replace fixed-space KNN with a learned embedding and deterministic continuous neighbors (ablation) | DnCNN image denoising · Urban100, σ=25 | PSNR: 30.08 -> 30.19 | Improved |
| use denoised rather than noisy patches for hard KNN matching (ablation) | DnCNN image denoising · Urban100, σ=25 | PSNR: 30.07 -> 30.08 | No clear effect |
| use stochastic Concrete relaxation instead of deterministic continuous neighbors (ablation) | N³Net image denoising · Urban100, σ=25 | PSNR: 30.19 -> 29.97 | Hurt |
| reduce the architecture from 3 × DnCNN blocks to 2 × DnCNN blocks (ablation) | N³Net image denoising · Urban100, σ=25 | PSNR: 30.19 -> 29.99 | Hurt |
| increase the selected-neighbor count up to four (ablation) | N³Net image denoising · Urban100, σ=25 and σ=50 | PSNR: k=1: 30.17; k=4: 30.27; k=7: 30.19; k=1: 26.76; k=4: 26.86; k=7: 26.82 | Mixed |
| augment DnCNN with an N³ block | DnCNN image denoising · Set12, BSD68, and Urban100 at σ=25, 50, and 70 | PSNR: Set12: 30.44 -> 30.55 (σ=25); 27.19 -> 27.43 (σ=50); 25.56 -> 25.90 (σ=70); BSD68: 29.23 -> 29.30 (σ=25); 26.23 -> 26.39 | Improved |
| augment VDSR with two N³ blocks | VDSR super-resolution · Set5 ×2, ×3, and ×4 | PSNR: 37.53 -> 37.57; 33.66 -> 33.84; 31.35 -> 31.50 | Improved |
| insert an N³ block after the sixth CNNet block | CNNet correspondence classification · St. Peter / St. Peter, 5° | MAP: 0.271 -> 0.316 | Improved |
| insert an N³ block after the sixth CNNet block | CNNet correspondence classification · St. Peter / Reichstag, 5° | MAP: 0.173 -> 0.231 | Improved |
| insert an N³ block after the sixth CNNet block | CNNet correspondence classification · Brown / Brown, 5° | MAP: 0.236 -> 0.293 | Improved |
| apply a post-hoc NN3D non-local step | DnCNN · BSD68, σ=50 | PSNR: -0.04 dB compared to DnCNN | Hurt |
Closest papers by meaning
- Non-Local Recurrent Network for Image Restoration (NeurIPS 2018 · similarity 0.86)
- On Convergence of Nearest Neighbor Classifiers over Feature Transformations (NeurIPS 2020 · similarity 0.85)
- Poly-NL: Linear Complexity Non-Local Layers With 3rd Order Polynomials (ICCV 2021 · similarity 0.85)
- NM-Net: Mining Reliable Neighbors for Robust Feature Correspondences (CVPR 2019 · similarity 0.85)
- The Neural Tangent Link Between CNN Denoisers and Non-Local Filters (CVPR 2021 · similarity 0.85)
- DNNR: Differential Nearest Neighbors Regression (ICML 2022 · 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.
