Research map / Architectures and efficiency / CNN design
Decoupled Dynamic Filter Networks
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 standard convolution with DDF | ResNet50 · ImageNet | Top-1 accuracy: 77.2 -> 79.1 | Helped |
| Replace standard convolution with DDF | ResNet101 · ImageNet | Top-1 accuracy: 78.9 -> 80.2 | Helped |
| Use DDF-Up instead of nearest-neighbor upsampling | Faster R-CNN with FPN · COCO minival | mAp: 37.4 -> 38.6 | Helped |
| Use DDF-Up instead of PAC-Net joint upsampling | Joint depth upsampling network · NYU Depth V2, 4× | RMSE: 2.39 -> 2.16 | Helped |
| Use DDF-Up instead of PAC-Net joint upsampling | Joint depth upsampling network · NYU Depth V2, 8× | RMSE: 4.59 -> 4.40 | Helped |
| Use DDF-Up instead of PAC-Net joint upsampling | Joint depth upsampling network · NYU Depth V2, 16× | RMSE: 8.09 -> 7.72 | Helped |
| Add spatial dynamic filters to channel-only DDF (ablation) | DDF-ResNet50 · ImageNet | Top-1 accuracy: 78.7 -> 79.1 | Helped |
| Add channel dynamic filters to spatial-only DDF (ablation) | DDF-ResNet50 · ImageNet | Top-1 accuracy: 74.4 -> 79.1 | Helped |
| Add spatial dynamic filters to the base convolution model (ablation) | ResNet50 · ImageNet | Top-1 accuracy: 77.2 -> 74.4 | Hurt |
| Replace filter normalization with Batch-Norm (ablation) | DDF-ResNet50 · ImageNet | Top-1 accuracy: 79.1 -> 76.0 | Hurt |
| Replace filter normalization with sigmoid activation (ablation) | DDF-ResNet50 · ImageNet | Top-1 accuracy: 79.1 -> 78.2 | Hurt |
| Increase squeeze ratio from 0.2 to 0.3 (ablation) | DDF-ResNet50 · ImageNet | Top-1 accuracy: 79.1 -> 79.0 | No clear effect |
Closest papers by meaning
- Dynamic Filter Networks (NeurIPS 2016 · similarity 0.90)
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.88)
- Decoupled Convolutions for CNNs (AAAI 2018 · similarity 0.87)
- DCFNet: Deep Neural Network with Decomposed Convolutional Filters (ICML 2018 · similarity 0.87)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 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.