basketball pose analysis
Detecting and analyzing human pose throughout freethrow motion
[Internship project for Acronis]
Goal: Build an application that analyzes basketball free throws from input smartphone video.
Motivation: We were approached by an NBA client for basketball analysis applications for both players and fans.
Method: There were three parts to this implementation
- Identifying human pose markers, basketball hoop, and basketball
- Classifying freethrow start and whether ball successfully went through hoop
- Estimating freethrow statistics (e.g. release angle, shot angle, bounce)
I fine-tuned Facebook’s Detectron2 model as well as trained our own model on an internal dataset of annotated ball and hoop images in MSCOCO format. I next developed several heuristics for classifying the success of the shot as well as player stats.
An interesting extension of this project would be mapping the video to a well-known player, or giving suggestions (the literature suggests that the release angle is highly correlated with the success of the shot).
Code: proprietary :’)