pytorchvideo

Why we included this project

Most video-understanding work stalls on plumbing before any modeling happens, and that is the gap PyTorchVideo targets. It packages pretrained architectures such as SlowFast and X3D, plus the video-specific data loaders, transforms, and evaluation utilities that are otherwise tedious to assemble by hand. This is a component library, not a finished application, aimed at action recognition, video classification, and temporal action detection. Because the blocks are PyTorch-native and hardware-accelerated, you can prototype quickly and stay close to research-grade baselines. The bundled benchmarks give you a solid reference point before you adapt models to your own footage.