Research map / Architectures and efficiency / CNN design
Scaling Up Your Kernels: Large Kernel Design in ConvNets Toward Universal Representations
Yiyuan Zhang, Xiaohan Ding, Xiangyu Yue
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 |
|---|---|---|---|
| remove identity shortcut when using 13x13 depth-wise kernels (ablation) | MobileNet V2 · ImageNet-1K | top-1 accuracy: 72.53 -> 53.98 | Hurt |
| replace 3x3 with 13x13 kernels when shortcuts exist (ablation) | MobileNet V2 · ImageNet-1K | top-1 accuracy: 71.76 -> 72.53 | Improved |
| add parallel 3x3 re-parameterization to 13x13 kernel (ablation) | MobileNet V2 · Cityscapes | mIoU: 75.67 -> 76.60 | Improved |
| add parallel 3x3 re-parameterization to 13x13 kernel (ablation) | MobileNet V2 · ImageNet-1K | top-1 accuracy: 72.53 -> 73.24 | Improved |
| enlarge last-stage kernels from 3x3 to 13x13 (ablation) | MobileNet V2 · Cityscapes | mIoU: 72.31 -> 74.62 | Improved |
| use Dilated Reparam Block instead of no re-parameterization (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 45.78 -> 46.37 | Improved |
| use Dilated Reparam instead of same-kernel-size parallel branches (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 46.07 -> 46.37 | Improved |
| add SE Block as extra depth structure (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 45.1 -> 46.5 | Improved |
| shrink kernels in stages 2-4 from 13 to 11 (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 46.5 -> 45.5 | Hurt |
| use 13x13 kernels in stage 1 instead of 3x3 (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 46.5 -> 44.9 | Hurt |
| add 3x3 DW conv to small-kernel blocks when scaling depth to 27 (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 47.8 -> 48.8 | Improved |
| use 27 large-kernel blocks instead of 9 large plus 18 small blocks (ablation) | vanilla large-kernel ConvNet · ADE20K | mIoU: 48.8 -> 49.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- InternImage-XL — InternImage: Exploring Large-Scale Vision Foundation Models With Deformable Convolutions (CVPR 2023) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.95)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.90)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.89)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.89)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.89)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.88)
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.
