Research map / Architectures and efficiency / CNN design
InternImage: Exploring Large-Scale Vision Foundation Models With Deformable Convolutions
Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, Yu Qiao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Add multi-group spatial aggregation (ablation) | InternImage-T / DCNv3 · ImageNet-1K val | top-1 accuracy: 82.3 -> 83.5 | Improved |
| Add multi-group spatial aggregation (ablation) | InternImage-T / DCNv3 · COCO val2017, Mask R-CNN 1× | box AP: 43.8 -> 47.2 | Improved |
| Normalize modulation scalars with softmax across sampling points (ablation) | InternImage-T / DCNv3 · ImageNet-1K val | top-1 accuracy: 65.7 -> 83.5 | Improved |
| Normalize modulation scalars with softmax across sampling points (ablation) | InternImage-T / DCNv3 · COCO val2017, Mask R-CNN 1× | box AP: 38.7 -> 47.2 | Improved |
| Share weights among convolution neurons (ablation) | InternImage-T / DCNv3 · ImageNet-1K val | top-1 accuracy: 83.6 -> 83.5 | No clear effect |
| Share weights among convolution neurons (ablation) | InternImage-T / DCNv3 · COCO val2017, Mask R-CNN 1× | box AP: 47.4 -> 47.2 | No clear effect |
| Increase DCNv3 kernel size from 3×3 to 5×5 (ablation) | InternImage-T · ImageNet-1K val | top-1 accuracy: 83.5 -> 83.6 | No clear effect |
| Increase DCNv3 kernel size from 3×3 to 7×7 (ablation) | InternImage-T · ImageNet-1K val | top-1 accuracy: 83.5 -> 82.8 | Hurt |
| Use scaling factors α=1.03 and β=1.40 instead of α=1.09 and β=1.36 (ablation) | InternImage-T scaled to about 100M parameters · ImageNet-1K val | top-1 accuracy: 84.9 -> 84.5 | Hurt |
| Use scaling factors α=1.12 and β=1.34 instead of α=1.09 and β=1.36 (ablation) | InternImage-T scaled to about 100M parameters · ImageNet-1K val | top-1 accuracy: 84.9 -> 83.1 | Hurt |
| Use a 3×3 dynamic sparse deformable convolution instead of dense large-kernel convolution (ablation) | InternImage · ImageNet-1K val, 224² | top-1 accuracy: 83.5 vs 82.8 | Improved |
Later papers that compared with this method
1 later accepted paper reports InternImage-XL as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Scaling Up Your Kernels: Large Kernel Design in ConvNets Toward Universal Representations (TPAMI 2025 · better on 2 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.87)
- Structured Initialization for Vision Transformers (NeurIPS 2025 · similarity 0.86)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.85)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.85)
- Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications (CVPR 2024 · similarity 0.84)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.84)
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.
