Research map / Architectures and efficiency / CNN design
Video Classification With Channel-Separated Convolutional Networks
Du Tran, Heng Wang, Lorenzo Torresani, Matt Feiszli
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 |
|---|---|---|---|
| channel-separate 3D convs (ir-CSN) in 26-layer net (ablation) | ResNet3D-26 · Kinetics-400 val (8-frame ablation) | video top-1: 65.3 -> 62.4 | Hurt |
| channel-separate with interaction preserved (ip-CSN) in 26-layer net (ablation) | ResNet3D-26 · Kinetics-400 val (8-frame ablation) | video top-1: 65.3 -> 64.6 | Hurt |
| channel separation (ir-CSN) in 50-layer net (ablation) | ResNet3D-50 · Kinetics-400 val (8-frame ablation) | video top-1: 69.4 -> 70.3 | Improved |
| channel separation (ip-CSN) in 50-layer net (ablation) | ResNet3D-50 · Kinetics-400 val (8-frame ablation) | video top-1: 69.4 -> 70.8 | Improved |
| channel separation (ir-CSN) in 101-layer net (ablation) | ResNet3D-101 · Kinetics-400 val (8-frame ablation) | video top-1: 70.6 -> 71.3 | Improved |
| channel separation (ip-CSN) in 101-layer net (ablation) | ResNet3D-101 · Kinetics-400 val (8-frame ablation) | video top-1: 70.6 -> 71.8 | Improved |
| preserve channel interactions (ip-CSN over ir-CSN) in 26-layer net (ablation) | CSN-26 · Kinetics-400 val (8-frame ablation) | video top-1: 62.4 -> 64.6 | Improved |
| remove ReLU between 1x1x1 and 3x3x3 layers (ablation) | ip-CSN · Kinetics-400 val (8-frame ablation) | video top-1 | No clear effect |
| add space-time decomposition of 3D filters (ablation) | ir-CSN-50 · Kinetics-400 val (8-frame ablation) | video top-1: 70.3 -> 69.7 | Hurt |
| apply group convolution to 1x1x1 layers (bottleneck-DG) (ablation) | bottleneck-D · Kinetics-400 val (8-frame ablation) | video top-1 accuracy | Hurt |
| pretrain on Sports1M before Kinetics (ablation) | ip-CSN-152 · Kinetics-400 val | video top-1: 77.8 -> 79.2 | Improved |
| pretrain on Sports1M before Kinetics (ablation) | ir-CSN-101 · Kinetics-400 val | video top-1: 76.2 -> 78.1 | Improved |
Closest papers by meaning
- Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks (ICCV 2019 · similarity 0.89)
- Dynamic Group Convolution for Accelerating Convolutional Neural Networks (ECCV 2020 · similarity 0.89)
- ClcNet: Improving the Efficiency of Convolutional Neural Network Using Channel Local Convolutions (CVPR 2018 · similarity 0.89)
- Exploiting Redundancy: Separable Group Convolutional Networks on Lie Groups (ICML 2022 · similarity 0.88)
- Interleaved Group Convolutions (ICCV 2017 · similarity 0.88)
- Unitary Convolutions for Learning on Graphs and Groups (NeurIPS 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.
