Research map / Video / Action recognition
In Defense of Image Pre-training for Spatiotemporal Recognition
Xianhang Li, Huiyu Wang, Chen Wei, Jieru Mei, Alan Yuille, Yuyin Zhou, Cihang Xie
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add ImageNet image pre-training (center-kernel init) (ablation) | ResNet50-3x3x3 on Kinetics-400 · Kinetics-400 (fixed budget) | top-1 accuracy: 70.7 -> 74.3 | Improved |
| add ImageNet image pre-training (ablation) | ResNet50-3x1x1 on Kinetics-400 · Kinetics-400 (fixed budget) | top-1 accuracy: 73.1 -> 74.9 | Improved |
| add STS convolution on top of image pre-training (ablation) | ResNet50-3x3x3 · Kinetics-400 (fixed budget) | top-1 accuracy: 74.3 -> 74.7 | Improved |
| add image pre-training and STS conv to SS-V2 models | Slowfast50 (4x16)-32f · Something-Something V2 | top-1 accuracy: 49.9 -> 57.2 | Improved |
| add STS conv on top of image pre-training (ablation) | Five 3D CNNs · Something-Something V2 | top-1 accuracy: +1.2% average | Improved |
| add image pre-training and STS conv to R2plus1D34-8f | R2plus1D34-8f · Kinetics-400 | top-1 accuracy: 68.7 -> 76.2 | Improved |
| add image pre-training and STS conv to X3D-S-13f | X3D-S-13f · Kinetics-400 | top-1 accuracy: 73.2 -> 73.5 | No clear effect |
| add ImageNet pre-training at 32 frames (ablation) | ResNet50-STS 3x3x3 · Kinetics-400 (32 frames) | top-1 accuracy: 74.4 -> 76.7 | Improved |
| pre-train on K400 frames instead of ImageNet (ablation) | ResNet50-STS 3x3x3 · Kinetics-400 (8 frames) | top-1 accuracy: 74.7 -> 73.4 | Hurt |
| pre-train on K400 frames instead of ImageNet (ablation) | X3D-S · Kinetics-400 (13 frames) | top-1 accuracy: 73.5 -> 72.8 | Hurt |
| use zero-init instead of inflation (ablation) | ResNet50-3x3x3 · Kinetics-400 | top-1 accuracy: 74.0 -> 74.3 | Improved |
| use zero-init instead of inflation (ablation) | ResNet50-3x1x1 · Something-Something V2 | top-1 accuracy: 61.8 -> 61.6 | Mixed |
Closest papers by meaning
- Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification (ECCV 2018 · similarity 0.91)
- Learning Spatio-Temporal Representation With Pseudo-3D Residual Networks (ICCV 2017 · similarity 0.90)
- Spatiotemporal Fusion in 3D CNNs: A Probabilistic View (CVPR 2020 · similarity 0.90)
- A Large-Scale Study on Unsupervised Spatiotemporal Representation Learning (CVPR 2021 · similarity 0.88)
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.88)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · 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.
