Advanced Basketball Statistics

Basketball has been one of my basic life-requirements since I learned how to coordinate my clumsy body. Although I was willing to compete all the time as I player, statistics were my main sport motivation, because it was truly fascinating how numbers could show what happened during a game. After a while, having realized that my ambitious goal of growing until 2 meters was not going to be fulfilled, I became a coach and learned something: it is impossible (for a player / ref / coach) to control everything that happens on court, not in real-time, but neither a posteriori when watching the game. Besides, classical statistics proved to be limited in terms on quantifying game-related aspects such as effort or defensive performance. This is the main reason of my life-time goal: provide basketball games with technological tools that could help to improve the performance (of a player / officials' decisions / the whole team).

Previously, I managed to combine this strong passion with studies, and I carried out some projects such as: a digital basketball coach board application (2011, high school), some Computer Vision algorithms that could detect a couple of classical basketball violations automatically (2015, Bachelor thesis) and a pre-trained Machine Learning module able to detect tactical patterns and classify different sets of plays (2017, Master thesis).

Deep Learning for Visual Recognition

During my Bachelor and Master studies I took several courses on classical Machine Learning, and while I was trying to recognize hand-written digits through feature extraction and simple Support Vector Machines, the ‘Big Data’ era emerged out of the blue. Although I was quite proud of my 0.95-number-accuracy model of uni labs, my Twitter timeline was full of topnotch creations such as food classifiers, music recommenders, artistic creations through neural style transfer or hilarious face-swapping algorithms... All with Deep Learning! Definitely, a hot topic since the GPU boom around 2014.

For me, at the very beginning, Deep Learning were black boxes that performed cool stuff, but that had nothing to do with my main interests. Nevertheless, literature proved 20-year-old Adrià wrong, and showed him many powerful possibilities in order to combine everything together. Convolutional Neural Networks and techniques such as Transfer Learning or Fine Tuning can be used for many Computer Vision techniques, such as detection and tracking of players, identification of their main biometrical keypoints, unique feature extraction, play clustering or forecasting. Companies in United States set complex sets of camera arrays in the ceiling of the arenas to obtain tracking data and send it to teams, but... Will it be possible to train similar models with single-camera systems and pose models?