Research map / Architectures and efficiency / CNN design
Dilated convolution with learnable spacings
Read the paperCheck 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 |
|---|---|---|---|
| Replace standard 3×3 convolutions with DCLS | ResNet-50 · ImageNet-1K val (crop 160) | Top-1 accuracy: 75.8 -> 76.5 | Helped |
| Use standard dilated convolution instead of standard convolution | ResNet-50 · ImageNet-1K val (crop 160) | Top-1 accuracy: 75.8 -> 71.7; 75.8 -> 70.1 | Hurt |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-T · ImageNet-1K val | Top-1 accuracy: 82.1 -> 82.5 | Helped |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-S · ImageNet-1K val | Top-1 accuracy: 83.1 -> 83.7 | Helped |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-B · ImageNet-1K val | Top-1 accuracy: 83.8 -> 84.1 | Helped |
| Use standard dilation rate 2 rather than DCLS | ConvNeXt-T · ImageNet-1K val | Top-1 accuracy: 82.1 -> 80.8 | Hurt |
| Use a dense 17×17 depthwise kernel rather than DCLS | ConvNeXt-T · ImageNet-1K val | Top-1 accuracy: 82.5 -> 82.0 | Hurt |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-T · ADE20K val, single-scale | mIoU: 46.0 -> 47.1 | Helped |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-S · ADE20K val, single-scale | mIoU: 48.7 -> 48.4 | Hurt |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-B · ADE20K val, single-scale | mIoU: 49.1 -> 49.3 | Helped |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt backbones · COCO val, Cascade Mask-RCNN 3× | Box AP: 50.4 -> 51.2; 51.9 -> 52.8; 52.7 -> 53.0 | Helped |
| Replace ConvNeXt depthwise convolutions with DCLS | ConvNeXt-S · ImageNet-Sketch | Accuracy: 37.1 -> 36.7 | Hurt |
Closest papers by meaning
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.87)
- Dilated Residual Networks (CVPR 2017 · similarity 0.86)
- Minimax optimality of convolutional neural networks for infinite dimensional input-output problems and separation from kernel methods (ICLR 2024 · similarity 0.86)
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.85)
- Frequency-Adaptive Dilated Convolution for Semantic Segmentation (CVPR 2024 · similarity 0.85)
- Receptive Fields As Experts in Convolutional Neural Architectures (ICML 2024 · 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.