Research map / Video / Action recognition
CAKES: Channel-wise Automatic KErnel Shrinking for Efficient 3D Networks
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 3D convolutions with channel-wise searched 1D/2D/3D kernels | C2FNAS · MSD test | Average DSC (%): 63.96 -> 64.42 | Helped |
| Replace 3D convolutions with channel-wise searched 1D/2D/3D kernels | C2FNAS · MSD test | Params (M): 22.50 -> 11.26 | Helped |
| Apply the architecture searched on MSD | C2FNAS · NIH Pancreas Segmentation | Average DSC: 84.59% -> 85.28% | Helped |
| Replace video 3D convolutions with CAKES P 1,2,3D | ResNet-50 · Something-Something V1, 16 frames | top1: 47.2 -> 49.4 | Helped |
| Use manual 2D kernels instead of standard 2D convolutions (ablation) | C2FNAS · MSD test | Average DSC (%): 61.09 -> 63.13 | Helped |
| Restrict the searched kernel set to 1D (ablation) | CAKES · MSD test | Average DSC (%): 63.96 -> 62.95 | Mixed |
| Restrict the searched kernel set to 2D (ablation) | CAKES · MSD test | Average DSC (%): 63.96 -> 64.28 | Helped |
| Use cost-priority rather than performance-priority search (ablation) | CAKES 1,2,3D · MSD test | Average DSC (%): 64.42 -> 64.15 | Hurt |
| Use cost-priority rather than performance-priority search (ablation) | CAKES 1,2,3D · NIH Pancreas Segmentation | Average DSC: 85.28% -> 84.85% | Hurt |
| Use manual 1D/2D assignments rather than searched assignments (ablation) | ResNet50 · Something-Something V1 | top1: 46.2 -> 47.1 | Hurt |
| Restrict candidate kernels to 1D/2D rather than 2D/3D (ablation) | CAKES performance-priority search · Something-Something V1 | top1: 47.4 -> 47.1 | Hurt |
| Expand the candidate set from 2D/3D to 1D/2D/3D under cost-priority search (ablation) | CAKES · Something-Something V1 | FLOPs: 41.4G -> 38.7G | Helped |
Closest papers by meaning
- LargeKernel3D: Scaling Up Kernels in 3D Sparse CNNs (CVPR 2023 · similarity 0.87)
- X3D: Expanding Architectures for Efficient Video Recognition (CVPR 2020 · similarity 0.86)
- Volumetric and Multi-View CNNs for Object Classification on 3D Data (CVPR 2016 · similarity 0.86)
- CT-Net: Channel Tensorization Network for Video Classification (ICLR 2021 · similarity 0.86)
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · similarity 0.85)
- OA-CNNs: Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation (CVPR 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.