Research map / Video / Video segmentation and tracking
A Multigrid Method for Efficiently Training Video Models
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 |
|---|---|---|---|
| Vary sampling grids and mini-batch size with long and short cycles | R50-SlowFast · Kinetics-400 | top-1 accuracy: 75.6 -> 76.4 | Helped |
| Vary sampling grids and mini-batch size with long and short cycles | R50-SlowFast · Kinetics-400 | speedup: 4.5$\times$ | Helped |
| Add a four-shape long cycle instead of using short cycle only (ablation) | Multigrid R50-SlowFast · Kinetics-400 | top-1 accuracy: 74.0% -> 76.4% | Helped |
| Add a three-shape short cycle instead of long-cycle only (ablation) | Multigrid R50-SlowFast · Kinetics-400 | top-1 accuracy: 74.5% -> 76.4% | Helped |
| Add the fine-tuning phase (ablation) | Default multigrid R50-SlowFast · Kinetics-400 | top-1 accuracy: 76.0% -> 76.4% | Helped |
| Apply multigrid training | R50-SlowFast · Something-Something V2 | top-1 accuracy: 60.9\mypm{0.31} -> 61.2\mypm{0.18} | Helped |
| Apply multigrid training | R50-SlowFast · Charades | mAP (%): 38.0\mypm{0.18} -> 38.2\mypm{0.06} | Helped |
| Apply multigrid training | I3D · Kinetics-400 | top-1 accuracy: 74.4 -> 74.8 | Helped |
| Apply multigrid training | I3D-NL · Kinetics-400 | top-1 accuracy: 75.5 -> 75.5 | No clear effect |
| Apply multigrid training | I3D · Kinetics-400, 1 GPU | training time (days): 6.7 -> 2.0 | Helped |
| Use a single long-cycle period instead of synchronizing cycles with learning-rate stages (ablation) | Multigrid R50-SlowFast · Kinetics-400 | top-1 accuracy: 76.4 -> 74.4 | Hurt |
| Use a cosine rather than stepwise learning-rate schedule (ablation) | Multigrid R50-SlowFast · Kinetics-400 | top-1 accuracy: 76.4 -> 75.3 | Hurt |
Closest papers by meaning
- Accelerating the Training of Video Super-resolution Models (AAAI 2023 · similarity 0.90)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.85)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.85)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.84)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.84)
- TrAct: Making First-layer Pre-Activations Trainable (NeurIPS 2024 · similarity 0.84)
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.