Research map / Video / Action recognition
Improving Video Model Transfer With Dynamic Representation Learning
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 |
|---|---|---|---|
| Pre-train with DRL using frozen clips | TSM ResNet-50 · HMDB-51 linear evaluation | accuracy: 54.64 -> 58.04 | Helped |
| Pre-train with DRL using frozen clips | TSM ResNet-50 · Diving-48 linear evaluation | accuracy: 17.01 -> 22.34 | Helped |
| Pre-train with DRL using frozen clips | TSM ResNet-50 · UCF-101 fine-tuning | accuracy: 95.14 -> 95.03 | No clear effect |
| Pre-train with DRL using frozen clips | TSM ResNet-18 · Jester 5-shot | 5-way accuracy: 39.07 -> 47.00 | Helped |
| Pre-train with DRL using frozen clips | 3D ResNet-18 · Mimetics domain generalization | Accuracy@1: 18.9 -> 26.4 | Helped |
| Use FGSM adversarial augmentation with ϵ = 8 (ablation) | 3D ResNet-18 on miniKinetics-200 · miniKinetics-200 | dynamic score (Distill): 0.33 -> 0.45 | Helped |
| Use PGD adversarial augmentation with ϵ = 8 (ablation) | 3D ResNet-18 on miniKinetics-200 · miniKinetics-200 | dynamic score (Distill): 0.33 -> 0.45 | Helped |
| Use direct min-max optimization with distillation and λ = 0.5 (ablation) | 3D ResNet-18 on miniKinetics-200 · miniKinetics-200 | dynamic score (Distill): 0.33 -> 0.61 | Helped |
| Use frozen clips for approximate min-max optimization with λ = 0.5 (ablation) | 3D ResNet-18 on miniKinetics-200 · miniKinetics-200 | dynamic score (Freeze): 1.06 -> 2.21 | Helped |
| Use shuffled clips for approximate min-max optimization with λ = 0.5 (ablation) | 3D ResNet-18 on miniKinetics-200 · miniKinetics-200 | dynamic score (Shuffle): 0.74 -> 1.23 | Helped |
| Use frame shuffling rather than frozen clips for approximate min-max optimization (ablation) | DRL linear evaluation · HMDB-51 | accuracy | Mixed |
| Increase the dynamic-loss weight λ with frozen-clip approximation (ablation) | DRL 3D ResNet-18 · HMDB-51 linear evaluation | accuracy: up to 5% | Helped |
Closest papers by meaning
- Dynamic Image Networks for Action Recognition (CVPR 2016 · similarity 0.89)
- Memory-Augmented Temporal Dynamic Learning for Action Recognition (AAAI 2019 · similarity 0.88)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.87)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · similarity 0.87)
- Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion (AAAI 2021 · similarity 0.87)
- Learning Spatio-Temporal Representation With Local and Global Diffusion (CVPR 2019 · similarity 0.87)
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.