We also observe that no single backbone performs best and that there is still a noticeable performance gap on certain generalization scenarios, motivating investigation into truly versatile backbones and better adaptation strategies.
[4/4]🧵

We systematically evaluate and compare the performance of leading Video Foundation Models (ViFMs) across a wide range of scientific domains within a unified framework. Our analysis shows that ViFMs provide a strong signal in highly diverse spatiotemporal applications.
[3/4]🧵

🧑‍💻 Inspect SciVid benchmark data and models with our colab demo: bit.ly/44BuJcS
Curious to evaluate your own video model on SciVid? Our code bit.ly/3Ge92YB provides data download instructions and supports evaluation of a HF 🤗 VideoMAE-B backbone, and is extensible to evaluate your own model.
[2/4]🧵