Research map / Video / Action recognition
Optimization Planning for 3D ConvNets
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 |
|---|---|---|---|
| Use optimization planning instead of the best hand-tuned schedule | DG-P3D · Kinetics-400 | top-1 accuracy: 78.0 -> 78.9 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · HMDB51 split 1 | top-1 accuracy: 57.7 -> 58.4 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · UCF101 split 1 | top-1 accuracy: 88.2 -> 89.1 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · ActivityNet | top-1 accuracy: 75.2 -> 76.5 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · SS-V1 | top-1 accuracy: 51.4 -> 52.8 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · SS-V2 | top-1 accuracy: 63.9 -> 65.5 | Helped |
| Use optimization planning rather than hand-tuning | DG-P3D · Kinetics-400 | top-1 accuracy: 78.0 -> 78.9 | Helped |
| Add the dual-head classifier to G-P3D (ablation) | G-P3D · Kinetics-400 | top-1 accuracy: 77.9 -> 78.9 | Helped |
| Insert an additional state in the optimization path (ablation) | DG-P3D optimization path · Kinetics-400 | top-1 accuracy: 0.2%~0.6% accuracy decrease | Hurt |
| Skip an intermediate state in the optimization path (ablation) | DG-P3D optimization path · Kinetics-400 | top-1 accuracy: 0.3%~1.0% accuracy decrease | Hurt |
| Use exponential rather than power curve fitting (ablation) | Knee-point estimation · 162 performance-epoch curves | RMSE: 1.010\times 10^{-3} -> 1.007\times 10^{-3} | Helped |
| Use different fitting functions for knee-point estimation (ablation) | Optimization planning · Training performance | final performance | No clear effect |
Closest papers by meaning
- 2D or not 2D? Adaptive 3D Convolution Selection for Efficient Video Recognition (CVPR 2021 · similarity 0.88)
- Spatial-Temporal Concept Based Explanation of 3D ConvNets (CVPR 2023 · similarity 0.86)
- Maximizing Spatio-Temporal Entropy of Deep 3D CNNs for Efficient Video Recognition (ICLR 2023 · similarity 0.86)
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.86)
- PointConv: Deep Convolutional Networks on 3D Point Clouds (CVPR 2019 · similarity 0.85)
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.85)
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.