Research map / Architectures and efficiency / CNN design
Dynamic Convolution: Attention Over Convolution Kernels
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 static convolution with dynamic convolution | MobileNetV2 ×1.0 · ImageNet val | Top-1 accuracy: 72.0 -> 75.2 | Helped |
| Replace static convolution with dynamic convolution | MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 65.4 -> 69.9 | Helped |
| Replace static convolution with dynamic convolution | MobileNetV3-Small · ImageNet val | Top-1 accuracy: 67.4 -> 70.3 | Helped |
| Use dynamic convolution in both backbone and head | MobileNetV2 ×0.5 keypoint model · COCO val2017 | AP: 59.2 -> 62.8 | Helped |
| Replace attention-weighted kernel aggregation with average aggregation (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 36.0 | Hurt |
| Replace weighted aggregation with maximum-attention kernel selection (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 0.1 | Hurt |
| Use attention weights from a different image (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 27.3 | Hurt |
| Restrict dynamic attention to high-level 14² and 7² resolutions (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 67.0 | Hurt |
| Use attention only at low-level 112², 56² and 28² resolutions (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 42.5 | Hurt |
| Use softmax temperature τ=1 instead of a large temperature (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 64.8 | Hurt |
| Anneal softmax temperature from τ=30 to τ=1 in the first 10 epochs (ablation) | DY-MobileNetV2 ×0.5 · ImageNet val | Top-1 accuracy: 69.4 -> 69.9 | Helped |
| Remove SE while retaining dynamic convolution (ablation) | DY-MobileNetV3-Small · ImageNet val | Top-1 accuracy: 70.3 -> 69.6 | Hurt |
Closest papers by meaning
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.92)
- Omni-Dimensional Dynamic Convolution (ICLR 2022 · similarity 0.91)
- KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters (NeurIPS 2026 · similarity 0.89)
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 2025 · similarity 0.87)
- Dynamic Convolutions Improve Transformers (NeurIPS 2026 · similarity 0.86)
- Dynamic Clustering Convolutional Neural Network (AAAI 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.