Deep Basketball Key Points Detection or Position Estimation

Abstract

This paper presents an approach for localizing basketball players and objects in real-world coordinates from videos. This is achieved by detecting 33 key points spread across the basketball court and using them as a reference to find the real-world coordinates of any pixel in the image that is known to be part of the ground plane. Moreover, a comprehensive synthetic dataset of basketball court images was generated using Blender: Each image was associated with a list of the 33 points’ 2D coordinates in the image. The key points are detected by a model performing two tasks: classification, determining whether a point is visible in the image, and regression, estimating the point’s precise location within the image. While the model was able to achieve more than 80% accuracy in detecting key points in the synthetic images, it was not able to generalize to real-world images, emphasizing the need to label some real-world data to fine-tune the model in the future.