Research map / Detection and segmentation / 3D detection and driving
LargeKernel3D: Scaling Up Kernels in 3D Sparse CNNs
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 plain sparse convolution blocks with SW-LK Conv | MinkowskiNet-34 · ScanNetv2 val | mIoU: 71.7 -> 73.5 | Helped |
| Replace plain sparse convolution blocks with SW-LK Conv | MinkowskiNet-34 · ScanNetv2 test | mIoU | Helped |
| Use 7×7×7 plain kernels instead of 3×3×3 (ablation) | MinkowskiNet-34 · ScanNetv2 val | mIoU: 71.7 -> 68.6 | Hurt |
| Add GELU activation (ablation) | MinkowskiNet-34 · ScanNetv2 | mIoU: 71.7 -> 70.2; 68.6 -> 68.4 | Hurt |
| Replace batch normalization with layer normalization (ablation) | MinkowskiNet-34 · ScanNetv2 | mIoU: 71.7 -> 66.8; 68.6 -> 65.0 | Hurt |
| Use depth-wise convolution (ablation) | MinkowskiNet-34 · ScanNetv2 | mIoU: 71.7 -> 70.6; 68.6 -> 68.7 | Hurt |
| Increase spatial group count from 2 to 16 (ablation) | MinkowskiNet-34 with 7×7×7 kernels · ScanNetv2 | mIoU: 68.8 -> 68.7 | Mixed |
| Use a 1×1×1 center group without shifting (ablation) | LargeKernel3D · ScanNetv2 | mIoU: 66.8 -> 73.5 | Helped |
| Increase center group size to 3×3×3 (ablation) | LargeKernel3D · ScanNetv2 | mIoU: 73.5 -> 70.6 | Hurt |
| Increase center group size to 5×5×5 (ablation) | LargeKernel3D · ScanNetv2 | mIoU: 73.5 -> 68.2 | Hurt |
| Add position embedding to 17×17×17 SW-LK (ablation) | CenterPoint · Waymo val vehicle LEVEL_1 | mAP: 71.5 -> 73.2 | Helped |
| Add position embedding to 17×17×17 SW-LK (ablation) | CenterPoint · Waymo val vehicle LEVEL_2 | mAP: 63.3 -> 65.1 | Helped |
Closest papers by meaning
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · similarity 0.90)
- To the Point: Efficient 3D Object Detection in the Range Image With Graph Convolution Kernels (CVPR 2021 · similarity 0.89)
- LinK: Linear Kernel for LiDAR-Based 3D Perception (CVPR 2023 · similarity 0.89)
- SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts (CVPR 2025 · similarity 0.86)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.86)
- OA-CNNs: Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation (CVPR 2024 · similarity 0.86)
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.