Research map / Video / Action recognition
Learning Spatio-Temporal Representation With Local and Global Diffusion
Check 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 local-global diffusion blocks to the TSN baseline | LGD-2D, ResNet-50 · Kinetics-600 validation | Top-1: 71.0 -> 72.5 | Helped |
| Add local-global diffusion blocks to the P3D baseline | LGD-3D, ResNet-50 · Kinetics-600 validation | Top-1: 71.2 -> 74.2 | Helped |
| Retain the input global feature when updating the global path (ablation) | LGD block · Kinetics-600 validation | Top-1: 71.6 -> 72.5; 72.7 -> 74.2 | Helped |
| Use global residual addition rather than channel-wise multiplication (ablation) | LGD block · Kinetics-600 validation | Top-1: 72.2 -> 72.5; 73.6 -> 74.2 | Helped |
| Add ImageNet pretraining (ablation) | LGD-2D, ResNet-50 · Kinetics-600 validation | Top-1: 72.5 -> 74.4 | Helped |
| Add ImageNet pretraining (ablation) | LGD-3D, ResNet-50 · Kinetics-600 validation | Top-1: 74.2 -> 75.8 | Helped |
| Add the local-global combination classifier (ablation) | LGD-2D, ResNet-50, ImageNet pretrained · Kinetics-600 validation | Top-1: 74.4 -> 74.8 | Helped |
| Add the local-global combination classifier (ablation) | LGD-3D, ResNet-50, ImageNet pretrained · Kinetics-600 validation | Top-1: 75.8 -> 76.3 | Helped |
| Train with 128-frame clips rather than 16-frame clips (ablation) | LGD-3D, ResNet-50, pretrained and combined · Kinetics-600 validation | Top-1: 76.3 -> 79.4 | Helped |
| Use ResNet-101 instead of ResNet-50 (ablation) | LGD-2D, pretrained and combined · Kinetics-600 validation | Top-1: 74.8 -> 76.7 | Helped |
| Use ResNet-101 instead of ResNet-50 (ablation) | LGD-3D, pretrained, combined, and 128-frame clips · Kinetics-600 validation | Top-1: 79.4 -> 81.5 | Helped |
| Fuse RGB and Flow predictions | LGD-3D, ResNet-101 · Kinetics-600 validation | Top-1: 81.5 -> 83.1 | Helped |
Closest papers by meaning
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.89)
- What Have We Learned From Deep Representations for Action Recognition? (CVPR 2018 · similarity 0.89)
- Spatiotemporal Residual Networks for Video Action Recognition (NeurIPS 2016 · similarity 0.88)
- Collaborative Spatiotemporal Feature Learning for Video Action Recognition (CVPR 2019 · similarity 0.88)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · similarity 0.88)
- StNet: Local and Global Spatial-Temporal Modeling for Action Recognition (AAAI 2019 · similarity 0.88)
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.