Follow these steps to create a video that tracks the location of a pickleball and obtain metrics on the model's performance.
Step 1 - Turn video into frames
Step 2 - Label frames by hand to get label csv
Step 3 - Adjust format of the label csv
Step 4 - Use label csv to turn frames into .npy files
Step 5 - Train model using .npy files and get new weights
Step 6 - Use video and new weights to get prediction csv
Step 7 - Use video and prediction csv to show ball trajectory
Step 8 - Use prediction csv and label csv to get model performance
For more information about each step, look into its respective folder.
GPU is required for training the model and getting predictions (steps 5 and 6). CPU can be used for everything else.
Kaggle CPU Docker Image
Kaggle GPU Docker Image
There is also a requirements.txt file straight from the Kaggle GPU image, but it may not be easily usable.
The weights are too big to upload to GitHub (>25MB), so I uploaded them to Google Drive. The old weights are from the 3-in-3-out model in the TrackNetV2 repository. The new weights are from our best model.
