Research map / Detection and segmentation / CNN features
Deformable Kernels: Adapting Effective Receptive Fields for Object Deformation
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 rigid 3×3 kernels with local DKs using scope 4×4 | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 76.3 -> 78.1 | Helped |
| Use local DKs with scope 3×3 instead of scope 4×4 (ablation) | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 78.1 -> 77.4 | Hurt |
| Increase local DK scope from 4×4 to 5×5 (ablation) | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 78.1 -> 77.8 | Hurt |
| Increase local DK scope from 4×4 to 9×9 (ablation) | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 78.1 -> 77.4 | Hurt |
| Use local instead of global DKs on 3×3 convolutions (ablation) | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 77.6 -> 78.1 | Helped |
| Add global DKs to 1×1 convolutions (ablation) | ResNet-50-DW with local 3×3 DKs · ImageNet 2012 val | top-1 accuracy (%): 78.1 -> 78.5 | Helped |
| Combine local DKs with Conditional Convolutions | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 77.6 -> 78.9 | Helped |
| Combine local DKs with Deformable Convolutions | ResNet-50-DW · ImageNet 2012 val | top-1 accuracy (%): 78.0 -> 79.0 | Helped |
| Replace rigid 3×3 kernels with local DKs using scope 4×4 | ResNet-50-DW Faster R-CNN + FPN · COCO test-dev | mAP: 36.6 -> 37.8 | Helped |
| Use local instead of global DKs on 3×3 convolutions (ablation) | ResNet-50-DW Faster R-CNN + FPN · COCO test-dev | mAP: 36.7 -> 37.8 | Helped |
| Add global DKs to 1×1 convolutions (ablation) | ResNet-50-DW Faster R-CNN + FPN with local 3×3 DKs · COCO test-dev | mAP: 37.8 -> 38.4 | Helped |
| Combine local DKs with Deformable Convolutions | ResNet-50-DW Faster R-CNN + FPN · COCO test-dev | mAP: 39.9 -> 40.6 | Helped |
Closest papers by meaning
- Deformable ConvNets V2: More Deformable, Better Results (CVPR 2019 · similarity 0.90)
- Image Deformation Meta-Networks for One-Shot Learning (CVPR 2019 · similarity 0.87)
- Extracting Deformation-Aware Local Features by Learning to Deform (NeurIPS 2021 · similarity 0.87)
- 4Deform: Neural Surface Deformation for Robust Shape Interpolation (CVPR 2025 · similarity 0.86)
- Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications (CVPR 2024 · similarity 0.85)
- Neural Pose Representation Learning for Generating and Transferring Non-Rigid Object Poses (NeurIPS 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.