VideoMAE

Why we included this project

VideoMAE is the official PyTorch implementation of a NeurIPS 2022 Spotlight paper that brought masked autoencoder pre-training to video understanding, and it remains a reference implementation for teams working on action recognition. If you are training models to classify or locate actions in video, say for sports analytics or media tagging, this repo shows how to pre-train a spatiotemporal Vision Transformer using only a small fraction of the labeled footage that fully supervised methods demand. The code supports multi-node distributed training, includes pretrained models and checkpoints for Kinetics-400, Something-Something, UCF101 and HMDB51, and offers a Colab notebook plus Hugging Face models so you can probe results without running heavy pipelines first. It is a research-oriented codebase rather than a plug-and-play service, so expect to adapt configs and data loaders to your own dataset. Teams that value data efficiency and want a proven starting point for video self-supervision will find it genuinely useful.