Facebook recently OSS’ed PyTorchVideo to facilitate video understanding research by providing model developers with reusable, modular, and efficient components. Designed for tasks like video classification, detection, and optical flow, the library also supports audio and text and includes packages for mobile hardware-specific optimizations and deployment. In addition, PyTorchVideo includes a model zoo with pretrained models, benchmarks, and data loaders and a toolkit of scripts for common video processing steps like decoding and tracking.