Research map / Architectures and efficiency / CNN design
Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs
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 |
|---|---|---|---|
| replace 3×3 depth-wise kernels with 13×13 kernels while retaining identity shortcuts (ablation) | MobileNet V2 · ImageNet | top-1 accuracy: 71.76 -> 72.53 | Helped |
| replace 3×3 depth-wise kernels with 13×13 kernels after removing identity shortcuts (ablation) | MobileNet V2 · ImageNet | top-1 accuracy: 68.67 -> 53.98 | Hurt |
| add parallel 3×3 re-parameterization to 9×9 kernels (ablation) | MobileNet V2 · ImageNet | top-1 accuracy: 72.67 -> 73.09 | Helped |
| add parallel 3×3 re-parameterization to 13×13 kernels (ablation) | MobileNet V2 · ImageNet | top-1 accuracy: 72.53 -> 73.24 | Helped |
| add parallel 3×3 re-parameterization to 13×13 kernels (ablation) | MobileNet V2 · Cityscapes | mIoU: 75.67 -> 76.60 | Helped |
| increase last-stage depth-wise kernel size from 3×3 to 13×13 (ablation) | MobileNet V2 · Cityscapes | mIoU: 72.31 -> 74.62 | Helped |
| increase stage kernels from 13-13-13-13 to 31-29-27-13 (ablation) | RepLKNet · ImageNet-1K | top-1 accuracy: 83.02 -> 83.07 | No clear effect |
| increase stage kernels from 13-13-13-13 to 31-29-27-13 (ablation) | RepLKNet · ADE20K | single-scale mIoU: 48.35 -> 49.17 | Helped |
| replace dense 9×9 depth-wise kernels with 3×3 depth-wise kernels at dilation 4 (ablation) | MobileNet V2 · ImageNet-1K | top-1 accuracy: 72.67 -> 57.23 | Hurt |
| replace dense 13×13 depth-wise kernels with 3×3 depth-wise kernels at dilation 6 (ablation) | MobileNet V2 · ImageNet-1K | top-1 accuracy: 72.53 -> 51.21 | Hurt |
| increase RepLKNet kernels from 3-3-3-3 to 31-29-27-13 (ablation) | RepLKNet · ADE20K | single-scale mIoU: 46.05 -> 49.17 | Helped |
| replace ConvNeXt 7×7 kernels with 31-29-27-13 kernels (ablation) | ConvNeXt-Tiny · ADE20K | single-scale mIoU: 44.6 -> 46.2 | Helped |
Closest papers by meaning
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.90)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.89)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.88)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.88)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.88)
- InceptionNeXt: When Inception Meets ConvNeXt (CVPR 2024 · similarity 0.87)
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.