Research map / Architectures and efficiency / CNN design
PSConv: Squeezing Feature Pyramid into One Compact Poly-Scale Convolutional Layer
Duo Li, Anbang Yao, Qifeng Chen
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 |
|---|---|---|---|
| replace standard 3x3 conv with PSConv in ResNet-50 | ResNet-50 · ImageNet val | top-1 error (%): 22.850 -> 21.126 | Improved |
| replace standard conv with PSConv in ResNeXt-101 (32x4d) | ResNeXt-101 · ImageNet val | top-1 error (%): 20.502 -> 19.498 | Improved |
| replace standard conv with PSConv in SE-ResNet-50 | SE-ResNet-50 · ImageNet val | top-1 error (%): 22.192 -> 20.814 | Improved |
| use cyclic interval t=4 instead of t=2 (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 21.948 -> 21.476 | Improved |
| increase cyclic interval from t=4 to t=8 (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 21.476 -> 21.634 | Hurt |
| add a second large dilation rate, {1,2,1,4} instead of {1,2,1,1} (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 22.368 -> 21.126 | Improved |
| enlarge single large dilation from 2 to 4, {1,4,1,1} vs {1,2,1,1} (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 22.368 -> 22.754 | Hurt |
| use aggressive dilation pattern {1,2,4,8} (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): over 5% drop | Hurt |
| add shift among filters (vary dilation along both axes vs input axis only) (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 21.658 -> 21.126 | Improved |
| vary dilation along both axes instead of output channel axis only (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 22.056 -> 21.126 | Improved |
| learnable binary mask for dilation 1 or 2 instead of static {1,2,1,4} (ablation) | PS-ResNet-50 · ImageNet val | top-1 error (%): 21.126 -> 21.138 | No clear effect |
| replace standard conv with PSConv in ResNeXt-29 (8x64d) | ResNeXt-29 · CIFAR-100 test | top-1 error (%): 18.074 -> 17.138 | Improved |
Closest papers by meaning
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.88)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.87)
- Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in CNNs (IJCV 2020 · similarity 0.87)
- Scale-Equalizing Pyramid Convolution for Object Detection (CVPR 2020 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.86)
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 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.
