Research map / Video / Action recognition
Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?
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 |
|---|---|---|---|
| Train ResNet-18 from scratch on Kinetics rather than the smaller video datasets | ResNet-18 · Kinetics vs UCF-101, HMDB-51, ActivityNet | validation accuracy: 40.1, 16.2, and 26.8% | Mixed |
| Increase ResNet depth from 18 to 152 layers (ablation) | 3D ResNet · Kinetics validation | Average accuracy (Top-1 and Top-5): 66.1 -> 73.7 | Helped |
| Increase ResNet depth from 152 to 200 layers (ablation) | 3D ResNet · Kinetics validation | Average accuracy (Top-1 and Top-5): 73.7 -> 73.7 | No clear effect |
| Use a wider network rather than ResNet-152 (ablation) | 3D ResNet · Kinetics validation | Average accuracy (Top-1 and Top-5): 73.7 -> 74.7 | Helped |
| Introduce group-convolution cardinality in ResNet (ablation) | 3D ResNet · Kinetics validation | Average accuracy (Top-1 and Top-5): 73.7 -> 75.4 | Helped |
| Replace ResNet-200 blocks with pre-activation blocks (ablation) | 3D ResNet-200 · Kinetics validation | Average accuracy (Top-1 and Top-5): 73.7 -> 73.4 | Hurt |
| Use DenseNet connections instead of ResNet-style connections (ablation) | 3D CNN · Kinetics validation | Average accuracy (Top-1 and Top-5): 73.7 vs 70.8; 72.3 | Hurt |
| Increase input clip length from 16 to 64 frames (ablation) | ResNeXt-101 · Kinetics test | Average accuracy (Top-1 and Top-5): 74.5 -> 78.4 | Helped |
| Pretrain ResNet-18 on Kinetics rather than training from scratch (ablation) | ResNet-18 · UCF-101 | Top-1 accuracy: 42.4 -> 84.4 | Helped |
| Pretrain ResNet-18 on Kinetics rather than training from scratch (ablation) | ResNet-18 · HMDB-51 | Top-1 accuracy: 17.1 -> 56.4 | Helped |
| Increase input clip length from 16 to 64 frames (ablation) | ResNeXt-101 · UCF-101 | Top-1 accuracy: 90.7 -> 94.5 | Helped |
| Increase input clip length from 16 to 64 frames (ablation) | ResNeXt-101 · HMDB-51 | Top-1 accuracy: 63.8 -> 70.2 | Helped |
Closest papers by meaning
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.91)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · similarity 0.90)
- Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset (CVPR 2017 · similarity 0.89)
- Spatiotemporal Fusion in 3D CNNs: A Probabilistic View (CVPR 2020 · similarity 0.89)
- Spatiotemporal Residual Networks for Video Action Recognition (NeurIPS 2016 · similarity 0.89)
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.89)
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.