Research map / Video / Action recognition
How Transferable are Video Representations Based on Synthetic Data?
Yo-whan Kim, Samarth Mishra, SouYoung Jin, Rameswar Panda, Hilde Kuehne, Leonid Karlinsky, Venkatesh Saligrama, Kate Saenko, Aude Oliva, Rogerio Feris
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 |
|---|---|---|---|
| pre-train on synthetic videos instead of Kinetics | TSN ResNet-50 · UCF101 and HMDB51 (high bias) fine-tuning | top-1 accuracy: 86.17 -> 83.40 (UCF101), 57.45 -> 54.38 (HMDB51) | Hurt |
| pre-train on synthetic videos instead of Kinetics | TSN ResNet-50 · Mini-SSV2, Diving48, IkeaFA, UAV-Human (low bias) fine-tuning | top-1 accuracy: 48.50 -> 49.69 (Mini-SSV2), 32.45 -> 35.57 (UAV-Human) | Improved |
| pre-train on synthetic videos instead of Kinetics | TSN ResNet-50 · UCF101 linear probing | top-1 accuracy: 69.92 -> 28.02 | Hurt |
| pre-train on synthetic videos instead of Kinetics | TimeSformer · UAV-Human fine-tuning | top-1 accuracy: 23.27 -> 24.95 | Improved |
| pre-train on synthetic videos instead of Kinetics | TimeSformer · Diving48 fine-tuning | top-1 accuracy: 46.37 -> 44.86 | Hurt |
| pre-train on synthetic videos instead of Kinetics | R(2+1)D ResNet-50 · UCF101 fine-tuning | top-1 accuracy: 87.21 -> 80.02 | Hurt |
| use 16 input frames instead of 8 for synthetic pre-training (ablation) | TSN synthetic · UAV-Human fine-tuning | top-1 accuracy: 35.57 -> 40.51 | Improved |
| increase number of pre-training classes (30, 90, 150) (ablation) | Kinetics and synthetic pre-training · UCF101, HMDB51, UAV-Human | top-1 accuracy | Improved |
| increase number of pre-training classes (30, 90, 150) (ablation) | Kinetics and synthetic pre-training · Mini-SSV2 and Diving48 | top-1 accuracy | No clear effect |
| increase samples per class (250, 750, 1000) (ablation) | Kinetics and synthetic pre-training · six downstream tasks | top-1 accuracy | No clear effect |
| replace Kinetics with Moments-in-Time real pre-training (ablation) | TimeSformer · UCF101 fine-tuning | top-1 accuracy: 92.09 -> 91.24 | No clear effect |
Closest papers by meaning
- Synthetic Humans for Action Recognition from Unseen Viewpoints (IJCV 2021 · similarity 0.89)
- Punching Bag vs. Punching Person: Motion Transferability in Videos (ICCV 2025 · similarity 0.88)
- Generating Human Action Videos by Coupling 3D Game Engines and Probabilistic Graphical Models (IJCV 2019 · similarity 0.88)
- Learning Video Representations of Human Motion From Synthetic Data (CVPR 2022 · similarity 0.87)
- Learning Human Action Recognition Representations Without Real Humans (NeurIPS 2023 · similarity 0.87)
- Pre-training for Action Recognition with Automatically Generated Fractal Datasets (IJCV 2025 · 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.
