Research map / Detection and segmentation / CNN features
Deformable ConvNets V2: More Deformable, Better Results
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 regular conv and pooling modules with DCNv1 modules | Faster R-CNN + ResNet-50 · COCO 2017 validation, shorter side 1000 | AP bbox: 34.7 -> 38.0 | Helped |
| Extend deformable conv from conv5 to conv3~conv5 (ablation) | Faster R-CNN + ResNet-50 · COCO 2017 validation, shorter side 1000 | AP bbox: 38.0 -> 40.4 | Helped |
| Add deformable RoIpooling to deformable conv in conv3~conv5 (ablation) | Faster R-CNN + ResNet-50 · COCO 2017 validation, shorter side 1000 | AP bbox: 40.4 -> 41.0 | Helped |
| Add modulation to deformable conv and RoIpooling (ablation) | Faster R-CNN + ResNet-50 · COCO 2017 validation, shorter side 1000 | AP bbox: 41.0 -> 41.7 | Helped |
| Add feature mimicking on foreground positive RoIs (ablation) | DCNv2 · COCO 2017 validation, Faster R-CNN | AP bbox: 41.7 -> 43.1 | Helped |
| Apply feature mimicking on foreground and background RoIs rather than none (ablation) | DCNv2 · COCO 2017 validation, Faster R-CNN | AP bbox: 41.7 -> 42.1 | Helped |
| Apply feature mimicking only on background RoIs rather than none (ablation) | DCNv2 · COCO 2017 validation, Faster R-CNN | AP bbox: 41.7 -> 41.7 | No clear effect |
| Add foreground-only feature mimicking without deformable layers (ablation) | Regular ConvNet · COCO 2017 validation, Faster R-CNN | AP bbox: 34.7 -> 35.0 | No clear effect |
| Pretrain DCNv2 offsets and modulation scalars on ImageNet (ablation) | DCNv2 with ResNet-101 · COCO 2017 test-dev detection | AP bbox: 44.8 -> 44.9 | No clear effect |
| Use multi-scale testing instead of single-scale shorter side 1000 (ablation) | DCNv2 Faster R-CNN + ResNet-101 · COCO 2017 test-dev | AP bbox: 44.8 -> 46.0 | Helped |
Closest papers by meaning
- Deformable Kernels: Adapting Effective Receptive Fields for Object Deformation (ICLR 2020 · similarity 0.90)
- Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications (CVPR 2024 · similarity 0.89)
- Deformable Part Region Learning for Object Detection (AAAI 2022 · similarity 0.88)
- Deformable Graph Convolutional Networks (AAAI 2022 · similarity 0.87)
- Deformable DETR: Deformable Transformers for End-to-End Object Detection (ICLR 2021 · similarity 0.87)
- Image Deformation Meta-Networks for One-Shot Learning (CVPR 2019 · similarity 0.87)
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.