DESIGNING A NEURAL NETWORK ARCHITECTURE TO ACCELERATE CAMERA CALIBRATION IN SOCCER MATCH ANALYTICS

Abstract

The article proposes a method to accelerate the camera calibration process in soccer match analytics while maintaining acceptable accuracy. The study focuses on modifying the High-Resolution Network (HRNet) architecture to reduce computational costs, making it suitable for real-time applications under limited hardware conditions. HRNet is a deep learning architecture known for maintaining high-resolution feature representations throughout its layers, which is particularly valuable for dense prediction tasks like keypoint detection and semantic segmentation. Unlike traditional models that downsample spatial information early, HRNet preserves detailed spatial features by processing multiple resolutions in parallel and continuously exchanging information across them.

The proposed by authors approach builds on HRNet’s strengths while improving its efficiency through applying three key strategies: simplifying network structures, applying knowledge distillation to transfer information from larger models to smaller ones, and adopting multi-task learning to handle keypoint and line detection within a single unified model.

The study evaluates several HRNet variants, including the standard simplified versions (W32, W18), as well as new architectures developed by the authors: an ultra-compact version (W6) and a multi-task model. The analysis focuses on the trade-off between speed and accuracy.. These models are trained and tested on the SoccerNet 2023 dataset, which offers a large and diverse set of annotated soccer images from multiple camera viewpoints. The evaluation uses practical metrics that reflect both calibration accuracy and completeness across varied match conditions.

Obtained results show that the developed W6 model achieves up to a 270% increase in processing speed compared to the original HRNet, with only a moderate drop in performance (12%). Meanwhile, the proposed multi-task architecture delivers the highest accuracy among the larger models trained for the same number of epochs, while also improving processing speed even in the smaller variants. Based on the obtained results conclusion is made that these compact and multi-task architectures offer a practical solution for fast, automated camera calibration in real-world sports analytics.