This project creates a model to identify specific actions during the gameplay of Australian rules football. The model predicts the action being completed during each frame from a video of gameplay such as the frames shown below.
The model identifies the following actions:
- mark
- kick
- handball
- tackle
Training of the model is completed in the notebook: 01_model_train
The project uses the R(2+1)D model architecture and a pre-trained model on 65 million social media videos (from Instagram) presented in Large-scale weakly-supervised pre-training for video action recognition (2019) paper, and a PyTorch implementation from Microsoft Computer Vision .