Research map / Architectures and efficiency / CNN design
Omni-Dimensional Dynamic Convolution
Read the paperCheck 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 convolutions with four-dimensional ODConv attention | MobileNetV2 and ResNet backbones · ImageNet validation | Top-1 Acc (%): 70.25 -> 73.97 | Helped |
| Use ODConv instead of competing dynamic convolution or attention modules | ResNet50 · ImageNet validation | Top-1 Acc (%): 77.34 -> 78.52 | Helped |
| Use ODConv (4×) in detector backbones | Faster R-CNN and Mask R-CNN · MS-COCO 2017 validation, AP | AP (%): 37.2 -> 39.2 | Helped |
| Increase the attention reduction ratio setting from r=1/16 to r=1/4 (ablation) | ODConv (1×), ResNet18 · ImageNet | Top-1 Acc (%): 73.10 -> 73.41 | Helped |
| Increase the number of convolutional kernels from 1 to 8 (ablation) | ODConv, ResNet18 · ImageNet | Top-1 Acc (%): 73.10 -> 74.08 | Helped |
| Add the spatial, input-channel, and filter attentions together (ablation) | ODConv, ResNet18 · ImageNet | Top-1 Acc (%): 72.42 -> 73.41 | Helped |
| Add kernel-wise attention to the three other attention types (ablation) | ODConv, ResNet18 · ImageNet | Top-1 Acc (%): 73.41 -> 74.33 | Helped |
| Remove temperature annealing (ablation) | ODConv (4×), ResNet18 · ImageNet | Top-1 Acc (%): 74.33 -> 73.29 | Hurt |
| Replace Sigmoid with Softmax for spatial attention (ablation) | ODConv (4×), ResNet18 · ImageNet | Top-1 Acc (%): 74.33 -> 73.97 | Hurt |
| Remove attention sharing across convolutional kernels (ablation) | ODConv (4×), ResNet18 · ImageNet | Top-1 Acc (%): 73.97 -> 74.16 | Helped |
| Apply ODConv to only selected convolutional layer groups rather than all but the first (ablation) | MobileNetV2 (0.5×) · ImageNet | Top-1 Acc (%): 68.26 -> 67.43 | Hurt |
| Preserve more spatial information by pooling attention inputs to 3×3 (ablation) | ODConv (1×), ResNet18 · ImageNet | Top-1 Acc (%): 73.41 -> 73.47 | Helped |
Closest papers by meaning
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.91)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.91)
- Frequency Dynamic Convolution for Dense Image Prediction (CVPR 2025 · similarity 0.89)
- CondConv: Conditionally Parameterized Convolutions for Efficient Inference (NeurIPS 2019 · similarity 0.89)
- An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures (ICML 2025 · similarity 0.85)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · 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.