3D Ball Localization From A Single Calibrated Image

Abstract

Ball 3D localization in team sports has various applications including automatic offside detection in soccer, or shot release localization in basketball. Today, this task is either resolved by using expensive multi-views setups, or by restricting the analysis to ballistic trajectories. In this work, we propose to address the task on a single image from a calibrated monocular camera by estimating ball diameter in pixels and use the knowledge of real ball diameter in meters. This approach is suitable for any game situation where the ball is (even partly) visible. To achieve this, we use a small neural network trained on image patches around candidates generated by a conventional ball detector. Besides predicting ball diameter, our network outputs the confidence of having a ball in the image patch. Validations on 3 basketball datasets reveals that our model gives remarkable predictions on ball 3D localization. In addition, through its confidence output, our model improves the detection rate by filtering the candidates produced by the detector. The contributions of this work are (i) the first model to address 3D ball localization on a single image, (ii) an effective method for ball 3D annotation from single calibrated images, (iii) a high quality 3D ball evaluation dataset annotated from a single viewpoint. In addition, the code to reproduce this research will be made freely available at https://github.com/gabriel-vanzandycke/ deepsport

Context-Aware 3D Object Localization from Single Calibrated Images: A Study of Basketballs

“…To accomplish this, we require a suitable dataset that provides high-resolution images with accurate ball annotations. While datasets such as Basket-APIDIS [34] and Soccer-ISSIA [7] have been previously utilized for related tasks, we exclude them from consideration as the concerns raised in reference [32] also apply to this work.…”

“…Ballistic dataset Beside evaluation of our model on the DeepSport dataset, in Section 4.6, we benchmark it also on the evaluation dataset introduced in Ref. [32] that features highly reliable 3D ball annotations from ballistic trajectories. This dataset is composed of 233 images of 2336 × 1756 pixels resolution, coming from 2 different basketball arenas.…”

“…Baseline As a baseline for our approach, we use a model introduced in Ref. [32] whereby the ball 3D position is obtained by estimating its diameter in the image space and combining knowledge of real ball size with calibration data.…”

“…The regression head, borrowed from Ref. [32], contains 3 layers, and outputs a scalar supervised to predict the ball height in pixels.…”

“…While 3D ball localization has been extensively studied for parabolic trajectories in calibrated videos [3,24,2,12,27,21,15,25], little attention has been given to the general challenge of 3D ball localization from a single calibrated image, leaving a significant gap in research. Recent works addressing this task involved the estimation of the ball's diameter in pixels and the reconstruction of its 3D location based on this estimation [32,17]. How- Predicted height ĥ is used to find the ball vertical projection on the ground floor (x, ỹ + ĥ) which can be projected to ( X, Ỹ , 0) using the calibration data.…”