Introduction:

The final project for my CS6140: Machine Learning class, I developed a pipeline to scrape human pose-estimation data from professional baseball pitching videos and apply it to various machine learning models to ultimately expose pitch-tipping phenomenon.  

​

What is pitch-tipping?

Pitch-tipping is the phenomenon when a pitcher’s subtle repetitive movements prior to the ball’s release ‘tip-off’ the hitter as to what type of pitch they are about to be thrown. Exposing pitch-tipping patterns are beneficial to the hitter and detrimental to the pitcher.

​

The process:

I implemented Google's MediaPipe platform to obtain human pose-estimation data from Major League Baseball's MLB Film Room dataset, a rich and intricately supervised open-sourced video library. After rigorous preprocessing, I applied the temporal data to various machine learning models including (3) logistic regressors, (3) support vector machines, and (3) deep neural networks. After fitting the models to the data per pitcher and determining the best model with its respective hyperparameters, I implemented unsupervised clustering algorithms such as t-SNE and UMAP to provide explainable insights. Using these insights in further plot-based analysis exposes biomechanical patterns in pitcher motions which the models likely leveraged in their training processes. ​​​