Breaking the Game: A Low-Cost System for 3D Human and Ball Pose Estimation in Basketball

Accurate 3D human pose estimation and ball tracking are fundamental to sports analytics, yet current state-of-the-art solutions often rely on expensive, proprietary systems with dense camera arrays, specialized hardware, and significant manual intervention. Moreover, these systems are rarely publicly available, making them difficult to reproduce or adapt in academic and low-resource settings. In this paper, we propose a low-cost, fully reproducible multiview system for 3D pose estimation of both players and the ball in basketball scenarios. Our setup employs off-the-shelf, static cameras distributed around a full-size court, supported by a robust calibration pipeline to enable triangulation-based 3D reconstruction. We leverage a custom YOLOv11 implementation for both ball detection and 2D human pose estimation across views, and 3D reconstruction through geometric methods without fine-tuning the model for the human pose task. To evaluate accuracy, we compare results against ground-truth optical motion capture data recorded in the same environment. Our experiments reveal the limitations of current pose models in distant-camera, real-game-like settings, underscoring the need for publicly available benchmarks in sports vision. To facilitate reproducible research, we provide a small annotated dataset of single-player basketball actions with a ball. Rather than focusing on full automation or commercial-level output, our goal is to highlight the gap between existing models and real-world sports conditions, and to offer a practical, open baseline for future method development and dataset expansion in unconstrained environments.