How dvanderhaar/uj-aqa-cricketvision works
This repository releases the CricketVision cricket-stroke dataset alongside research code for action-quality assessment. The documented workflow combines video-frame and pose-keypoint inputs, learns representations with a two-stream I3D-AE-LSTM autoencoder, and regresses body-part quality scores. Dataset clips include phase annotations. The source excerpts were unavailable, so the model stages and their links below reflect README-level documentation, not inspected notebook internals.
Main entry points: ViT_I3D.ipynb Two-Stream Autoencoder · Videos.zip Stroke Clips · JSON.zip Phase Annotations · README.md AQA Benchmark
Components 10
CricketVision Dataset
- Stroke Clips
CricketVision_dataset_release/Videos.zip - Phase Annotations
CricketVision_dataset_release/JSON.zip - Annotation Structure
CricketVision_dataset_release/Annotation Structure.txt - Dataset Attributes
Dataset Attributes.md
Assessment Model
- Video Frames
I3D_AE_LSTM _Code_Release/ViT_I3D.ipynb - Pose Keypoints
I3D_AE_LSTM _Code_Release/ViT_I3D.ipynb - Two-Stream Autoencoder
I3D_AE_LSTM _Code_Release/ViT_I3D.ipynb - MLP Regressor
I3D_AE_LSTM _Code_Release/ViT_I3D.ipynb
Evaluation
- Body-Part Scores
I3D_AE_LSTM _Code_Release/ViT_I3D.ipynb - AQA Benchmark
README.md
Connections 7
- Stroke Clips → Phase Annotations: annotated with
README.md - Stroke Clips → Video Frames: provides video
README.md - Video Frames → Two-Stream Autoencoder: input to
README.md - Pose Keypoints → Two-Stream Autoencoder: input to
README.md - Two-Stream Autoencoder → MLP Regressor: features to
README.md - MLP Regressor → Body-Part Scores: predicts
README.md - Body-Part Scores → AQA Benchmark: evaluated by
README.md