Jules White, associate dean for strategic learning programs and associate professor of computer science and computer engineering, and Carlos Olea, a Ph.D. student in the Department of Computer Science, developed an AI software called a temporal relational network to help determine the context and mechanics behind each shot a player takes.
"I’m really excited about the potential for AI to help amateurs at home learn and improve," White told Fox News Digital. "I think there’s a limited amount of time and money that many people have to get access to coaching, so I think it expands the access to information that you need to improve."
The idea came about after a former student connected White and Olea with NOAH Basketball, a company that uses cameras and cutting-edge software to watch practices and gather shooting statistics for NBA and NCAA players. Using facial recognition and computer vision, NOAH generates detailed statistics about how and where somebody is shooting, the trajectory of the ball, and where they missed shots from.
NOAH provided more than 50,000 hours of video footage, which the researchers used to classify five different shot types: Free throw, catch and shoot, off the dribble, jab step fake, and step back shot. Using the AI, they were able to optimize shot type recognition, achieving an accuracy of 96.8% on 1,500 novel shots.
White said the researchers were interested in knowing the statistics but also the context behind the stats, like if the player was just standing there taking shots over and over, compared to if they were in the middle of the game, or had just received a pass.
"We connected and talked through having a research project where we would look into developing deep learning models to help them identify what was going on," White said. "Basically the AI is trying to learn and understand what are the movements that relate to different types of shots. We’re basically teaching it what things look like on video."
White said doing that is a human job in basketball, as people in the sport have to watch these thousands of videos and individually label each one by the type of shot.
He said Olea's role was using all of those examples to teach the AI how to identify the shot type, with the researcher's code essentially telling the AI what it did wrong, and where it made mistakes. White called it an interactive process where the AI is given examples and tries to predict what the correct answer is for those examples. After that, the program goes back and sees what it did wrong, and it updates its understanding of those movements.
"The AI that we designed can be thought of as a very naïve person trying to learn something, in this case how to identify those different types of basketball shots," Olea said.
White said you can probably gain the same type of information by following a player around with a clipboard, though he noted it would be "enormously expensive to do that."
"For some players, you can justify that expense, but for everybody else, this can automate that process of collecting that information that the player and coach can use to make decisions on how they can practice," he said.
Olea said that simply categorizing shots is one small portion of basketball, as there are other factors that can impact the shot, like how good the defender is, or how fast the player is.
The researchers said they started with basic things, but they hope to grow – with more research – to other portions, eventually assigning attributes to different players. They said that being partnered with NOAH was very helpful in both understanding the statistics and context behind each shot.
"We started with these first five things (the five shot types), but to be effective from a research perspective you kind of need a partner like that who really understands what matters and what are the statistics you want to know and what are the contextual things," White said. "So going forward, you would want exactly things like that."
Over time, White said he is hopeful that the sports world will see more of this type of research. He believes that companies, like NOAH, could eventually take this type of approach and apply it to real-world products and commercial systems that you can go and use.
"So my bet is that over time the type of advanced things that we are talking about here could be done on your phone," he added, saying it could be an app installed on your phone, that people can use at home or at the gym. "I think the long-term vision as the AI gets more sophisticated is that it would head in a direction like that."
"That’s kind of where it gets exciting is taking just something very raw that people are collecting now, like videos of themselves doing something … and giving you that extra bit of information to help you improve," he added. "So that’s where I would love to have something like that for all sporting-related things."
Olea who grew up spending hours practicing baseball, sometimes by himself, said having that technology would have been very useful during his youth.
"Having this type of feedback would be immensely useful for improving my own play and just understanding what I need to do to improve as a player," he added.
White and Olea talked about the shooting analysis in an award-winning conference paper titled, "Analysis of Deep Learning Action Recognition for Basketball Shot Type Identification."