Research map / Architectures and efficiency / CNN design
Adaptive Convolutions With Per-Pixel Dynamic Filter Atom
Ze Wang, Zichen Miao, Jun Hu, Qiang Qiu
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 both conv layers with ACDA | LeNet · CIFAR-10 | top-1 error: 24.74 -> 16.27 | Improved |
| replace both conv layers with ACDA | LeNet · CIFAR-100 | top-1 error: 56.60 -> 49.53 | Improved |
| use ACDA blocks instead of standard convs, with far fewer parameters | ResNet-18 · ImageNet | top-1 error: 30.24 -> 28.81 | Improved |
| use per-pixel ACDA instead of standard convs | crowd counting baseline · SHTech-A | MAE: 67.73 -> 57.88 | Improved |
| use per-pixel ACDA instead of standard convs | crowd counting baseline · UCF-QNRF | MAE: 124.77 -> 99.22 | Improved |
| per-pixel ACDA instead of per-image CondConv-l | crowd counting · SHTech-A | MAE: 63.82 -> 56.04 | Improved |
| use ACDA instead of standard convs | denoising network (MSE loss) · SIDD | PSNR: 38.74 -> 38.96 | Improved |
| train ACDA with VDNet variational framework instead of MSE | ACDA · SIDD | PSNR: 38.96 -> 39.32 | Improved |
| use two-layer implementation instead of direct per-pixel DFN | per-pixel adaptive conv · ImageNet training memory | memory: OOM -> 4.6GB x 4 | Improved |
| use ACDA two-layer implementation instead of standard conv | Ad-ResNet-s · ImageNet training | time per iteration: 0.53s -> 0.46s | Improved |
| increase number of atoms m from 3 to 7 (ablation) | ACDA · RealSR Final x3 | PSNR: 30.52 -> 30.73 | Improved |
| increase number of atoms m from 7 to 11 (ablation) | ACDA · RealSR Final x3 | PSNR: 30.73 -> 30.75 | No clear effect |
Closest papers by meaning
- Dynamic Filter Networks (NeurIPS 2016 · similarity 0.89)
- Large Convolutional Model Tuning via Filter Subspace (ICLR 2025 · similarity 0.88)
- Decoupled Dynamic Filter Networks (CVPR 2021 · similarity 0.88)
- Bi-volution: A Static and Dynamic Coupled Filter (AAAI 2022 · similarity 0.88)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.86)
- An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures (ICML 2025 · 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.
