Hoopai

Predicting basketball shot probability using modern computer vision technology. The goal of this project is to first develop a tool similar to HomeCourt and expand on shot predictions instead of just shot tracking.

Environment

Python 3.8.18

Models and Dataset

The datasets used to train are located:

• Here...

The trained models can be found

• Here...

Tools

Used labelmestudio.

Running

Put video in ...

python main.py --input_video './videos/indoor2.mp4' --save_frames True --output_dir ./output/run1 --show_stats True   

Immediate Todos

Eventual Todos

• Set up simple LSTM RNN for shot prediction based on keypoints.
• User frontend built using React to allow users to use.
• Build mobile application
• Basketball shot arc prediction: (1) Linear regression (2) Physics?
• Improved method to handle frames where ball is not detected?

Development

2023.12.31 V2

• Improved ball and rim detection with bigger dataset.
• Simple state (holding, shot, score) recognition using overlap of bounding boxes (ball on person, ball on rim).
• Add some UI support for displaying detection

2023.12.08 V1:

• Proof of concept with Yolov8 for pose detection, ball detection and Rim Detection
• Strictly python backend. Run through video frame by frame with CV2. Run inference using trained models.
• Use two seperate models, one to perform pose detection, one to perform ball and rim detection