Research map / Detection and segmentation / CNN features
Non-Local Neural 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 |
|---|---|---|---|
| Add 5 non-local blocks to I3D (ablation) | I3D ResNet-101 · Kinetics val | top-1 accuracy: 74.4 -> 76.0 | Helped |
| Add 5 non-local blocks to I3D | I3D ResNet-101 · Charades trainval/test | mAP: 37.2 -> 39.5 | Helped |
| Add 1 non-local block to the Mask R-CNN backbone | Mask R-CNN ResNet-101 · COCO val2017 | box AP: 39.5 -> 40.8 | Helped |
| Add 4 non-local blocks to the keypoint head | Mask R-CNN ResNet-101 · COCO val2017 | keypoint AP: 65.1 -> 66.0 | Helped |
| Add 1 non-local block to the backbone, in addition to 4 in the keypoint head | Mask R-CNN ResNet-101 · COCO val2017 | keypoint AP: 66.0 -> 66.5 | Helped |
| Use Gaussian, embedded Gaussian, dot-product, or concatenation affinity instead of the other tested forms (ablation) | C2D ResNet-50 with one non-local block · Kinetics val | top-1 accuracy: 72.5 to 72.9 | No clear effect |
| Place the block in res5 rather than res2, res3, or res4 (ablation) | C2D ResNet-50 with one non-local block · Kinetics val | top-1 accuracy: 72.9 -> 72.3 | Hurt |
| Allow non-local interactions across spacetime rather than space only (ablation) | C2D ResNet-50 with 5 non-local blocks · Kinetics val | top-1 accuracy: 72.9 -> 73.8 | Helped |
| Allow non-local interactions across spacetime rather than time only (ablation) | C2D ResNet-50 with 5 non-local blocks · Kinetics val | top-1 accuracy: 73.1 -> 73.8 | Helped |
| Increase the number of non-local blocks from 1 to 10 (ablation) | C2D ResNet-50 · Kinetics val | top-1 accuracy: 72.7 -> 74.3 | Helped |
| Replace I3D with a 5-block non-local C2D model (ablation) | ResNet-101 video classifier · Kinetics val | top-1 accuracy: 74.4 -> 75.1 | Helped |
| Use 128-frame clips rather than 32-frame clips (ablation) | Kinetics video classifiers · Kinetics val | top-1 accuracy | Helped |
Closest papers by meaning
- Compact Generalized Non-local Network (NeurIPS 2018 · similarity 0.90)
- Few-Shot Learning With Localization in Realistic Settings (CVPR 2019 · similarity 0.86)
- Non-Local Recurrent Network for Image Restoration (NeurIPS 2018 · similarity 0.86)
- Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs (ICLR 2024 · similarity 0.84)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.84)
- Recognize Any Regions (NeurIPS 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.