Research map / Architectures and efficiency / CNN design
Convolution with even-sized kernels and symmetric padding
Shuang Wu, Guanrui Wang, Pei Tang, Feng Chen, Luping Shi
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 |
|---|---|---|---|
| Use symmetric padding with 2×2 kernels instead of asymmetric padding (ablation) | ResNet-50 0.5× · ImageNet | top-1 error (%): 31.0 -> 28.4 | Improved |
| Use 2×2 symmetric-padding kernels instead of 3×3 kernels | ResNet-50 · ImageNet | top-1 error (%): 23.8 -> 24.0 | Hurt |
| Use 2×2 symmetric-padding kernels instead of 3×3 kernels | DenseNet-121 · ImageNet | top-1 error (%): 24.6 -> 25.1 | Hurt |
| Use C2sp blocks instead of ShiftNet blocks | Compact CNNs · CIFAR100, 56- and 110-layer models | error (with mixup, %): 21.31 -> 20.91; 20.47 -> 19.52 | Mixed |
| Use C2sp blocks instead of depthwise-separable convolution blocks | Compact CNNs · CIFAR100 | memory (MB): 1707 -> 1219; 3170 -> 1951 | Mixed |
| Add cutout and mixup augmentation | Wide-DenseNet C2sp · CIFAR10 test | error (%): 3.54 -> 2.44 | Improved |
| Replace 3×3 kernels with C2sp | Wide-DenseNet · CIFAR10 test | error (%): 3.81 -> 3.54 | Improved |
| Use C2sp instead of asymmetric C2 kernels (ablation) | GAN generator · CIFAR10 generation | FID: non-convergence -> 23.35$\pm$0.26 | Improved |
| Use C4sp instead of asymmetric C4 kernels (ablation) | GAN generator · LSUN bedroom generation | FID: 39.17$\pm$5.52 -> 24.61$\pm$2.45 | Improved |
| Use C4sp instead of asymmetric C4 kernels (ablation) | GAN generator · CelebA-HQ generation | FID: 37.93$\pm$7.54 -> 30.22$\pm$3.02 | Improved |
| Replace C3 kernels with C2sp | GAN generator · CIFAR10 generation | FID: 24.12$\pm$0.47 -> 23.35$\pm$0.26 | Improved |
| Replace C3 kernels with C4sp | GAN generator · LSUN bedroom and CelebA-HQ generation | FID: 36.04$\pm$7.10 -> 24.61$\pm$2.45; 43.39$\pm$5.78 -> 30.22$\pm$3.02 | Improved |
Closest papers by meaning
- Scaling Up Your Kernels: Large Kernel Design in ConvNets Toward Universal Representations (TPAMI 2025 · similarity 0.87)
- Approximation and Learning with Deep Convolutional Models: a Kernel Perspective (ICLR 2022 · similarity 0.87)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.86)
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.86)
- The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods (ICLR 2021 · similarity 0.86)
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.86)
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.
