Research map / Architectures and efficiency / CNN design
Spatially-Adaptive Gradient Re-parameterization for 3D Large Kernel Optimization
Ho Hin Lee, Quan Liu, Shunxing Bao, Yuankai Huo, Bennett Allan Landman
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 LRBM learnable spatial bias to 3D UX-Net (k=7) (ablation) | 3D UX-Net (k=7) · AMOS CT | average Dice: 0.890 -> 0.897 | Improved |
| use learnable LRBM generator instead of fixed distance prior (ablation) | Rep3D · AMOS CT (from scratch) | average Dice: 0.902 -> 0.910 | Improved |
| use learnable LRBM generator instead of fixed distance prior (ablation) | Rep3D · AMOS MRI (from scratch) | average Dice: 0.855 -> 0.864 | Improved |
| use learnable LRBM generator instead of fixed distance prior (ablation) | Rep3D · KiTS | mean Dice: 0.727 -> 0.736 | Improved |
| add spatial prior modulation (fixed prior) over vanilla plain large-kernel training (ablation) | Rep3D · AMOS (60k iterations) | validation Dice: 0.892 -> 0.902 | Improved |
| add learnable LRBM over vanilla plain large-kernel training (ablation) | Rep3D · AMOS (60k iterations) | validation Dice: 0.892 -> 0.910 | Improved |
| add learnable LRBM, training time cost (ablation) | Rep3D · AMOS | training time (hrs): 17.3 -> 17.5 | No clear effect |
| use 3-layer instead of 2-layer generator (ablation) | Rep3D LRBM · AMOS CT | average Dice: 0.910 -> 0.899 | Hurt |
| use 1-layer instead of 2-layer generator (ablation) | Rep3D LRBM · AMOS CT | average Dice: 0.904 -> 0.910 | Hurt |
| increase generator kernel from 1x1x1 to 7x7x7 (ablation) | Rep3D LRBM generator · AMOS CT | average Dice: 0.905 -> 0.910 | Improved |
| generator kernel 5x5x5 instead of 3x3x3 (ablation) | Rep3D LRBM generator · AMOS CT | average Dice: 0.906 -> 0.903 | No clear effect |
| add a parallel 3x3x3 branch with higher learning rate (AdamW) (ablation) | Rep3D 21x21x21 CSLA · AMOS | mean Dice: 0.862 -> 0.889 | Improved |
Closest papers by meaning
- LinK: Linear Kernel for LiDAR-Based 3D Perception (CVPR 2023 · similarity 0.87)
- LargeKernel3D: Scaling Up Kernels in 3D Sparse CNNs (CVPR 2023 · similarity 0.87)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.86)
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · similarity 0.86)
- Enhancing High-Resolution 3D Generation through Pixel-wise Gradient Clipping (ICLR 2024 · similarity 0.85)
- WildFusion: Learning 3D-Aware Latent Diffusion Models in View Space (ICLR 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.
