Research map / Architectures and efficiency / CNN design
Learning Sparse High Dimensional Filters: Image Filtering, Dense CRFs and Bilateral Neural Networks
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learn general bilateral filter weights instead of fixing a Gaussian kernel | Joint bilateral color upsampling · Color upsampling, Pascal VOC12 and 2MP test images, 2x–16x | PSNR: 33.46 / 37.93 -> 34.05 / 38.74; 31.87 / 35.66 -> 32.28 / 36.38; 30.51 / 33.92 -> 30.81 / 34.41; 29.19 / 32.24 -> 29.52 / | Helped |
| Replace bicubic interpolation with learned bilateral upsampling | Depth prediction CNN · Depth upsampling, NYU test | RMSE: 0.753 -> 0.748 | Helped |
| Learn the dense CRF pairwise kernel rather than using a Gaussian kernel | Semantic segmentation with DeepLab and mean-field inference · Pascal VOC12 validation/test, 1-step and 2-step mean field | IoU change: +2.48 -> +2.93; +3.38 -> +3.71 | Helped |
| Learn the dense CRF pairwise kernel | Material segmentation CNN · MINC test, pixel accuracy | Pixel accuracy change: +7.91 / +6.28 -> +9.48 / +6.23; +9.68 / +7.35 -> +11.89 / +6.93 | Helped |
| Learn the dense CRF pairwise kernel | Material segmentation CNN · MINC test, class-averaged pixel accuracy | Class-averaged pixel accuracy change: +6.28 -> +6.23; +7.35 -> +6.93 | Mixed |
| Give separate kernels to the first and second mean-field steps (ablation) | Learned CRF for semantic segmentation · Pascal VOC12 validation/test, 2-step mean field | IoU change: +3.71 -> +3.85 / +3.37 | Helped |
| Replace the first spatial convolution with a bilateral convolution | LeNet and DeepCNet · Assamese character recognition, test | Accuracy: 69.10 -> 74.98; 80.02 -> 84.21 | Helped |
| Replace the first spatial convolution with bilateral convolution on original-resolution MNIST | LeNet · MNIST, original training and test data | Accuracy: 0.9919 -> 0.9903 | Hurt |
| Replace the first spatial convolution with bilateral convolution when training on subsampled MNIST at 100% and 60% | LeNet · MNIST, subsampled training and test data | Accuracy: 0.9660 -> 0.9844; 0.9348 -> 0.9534 | Helped |
| Replace the first spatial convolution with bilateral convolution when training on subsampled MNIST at 20% | LeNet · MNIST, 20% training and test samples | Accuracy: 0.8928 -> 0.9042 | Helped |
| Use a learned bilateral filter rather than a Gaussian bilateral filter | Single-filter image denoising · BSDS500 test | PSNR: $26.51$ -> $26.58$ | Helped |
| Add a spatial filter to the learned bilateral filter (ablation) | Single-filter image denoising · BSDS500 test | PSNR: $26.58$ -> $26.65$ | Helped |
Closest papers by meaning
- Provable Filter Pruning for Efficient Neural Networks (ICLR 2020 · similarity 0.87)
- Compressing Convolutional Neural Networks via Factorized Convolutional Filters (CVPR 2019 · similarity 0.87)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.87)
- Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018 · similarity 0.86)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.85)
- Deep Rank-One Tensor Functional Factorization for Multi-Dimensional Data Recovery (AAAI 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.