ulzee / pitch_type

pitch_type

• Pitch type classification from videos of baseball games

1. Find hyperparameters for RNN

• hyperparameters
• Trained differen convolutional and recurrent NN models
• Tested hyperparameters systematically with csv file and with genetic programming (but tended to make only fully connected layers)
• Best accuracy 64% on all data

2. Explore data:

• see plots of data here
• Coordinate trajectories
• Investigate data by plotting mean and different examples of joints by pitch type
• Testing different interpolation/smoothing
• Filling in missing values does not work properly
• Tried different filters (Kalmann, Gaussian, Cubic and linear interpolation) to fill in the values and smoothen the curve
• Problem: Sometimes the outliers are actually the right values
• Used LSTM to learn coordinate trajectories to fill in missing values, works but not plotted yet
• ML coord fill in RNN here

3. Process videos to joint coordinates

• see Pose estimation
• Aim: in the end one system for real time inferences from videos
• stitched together Estelle's preprocessing and my model
• Problem with tensorflow pytorch compatability
• Evaluated times for different steps

4. RNN for getting pitch type directly from video

• maybe information loss when using coordinates and not videos
• trained LSTM on video arrays directly
• see this folder
• see notebook for exploration
• Maybe loss of information when getting the coordinates - therefore trained LSTM directly on videos
• up to 50% Acc
• possible todo: try 3D conv net