GloFANet: Global Feature Augmented Networks for Data-Efficient Soccer Field Keypoint Detection

Keypoint detection in soccer broadcasts is crucial for tasks such as player positioning and goal event detection. Traditional algorithms like SIFT, which ensure scale and rotation invariance, and the Hough Line Transform, effective for straight line detection, lack adaptability to perspective changes and occlusion. While deep learning models adapt better, they still lose pixel-level details needed for precise keypoint localization in complex scenes like soccer fields and sports courts. To address these issues, the High-Resolution Network has been introduced to preserve pixel-level details through multi-scale feature fusion, but it still faces significant challenges under limited data conditions. To overcome these problems, we propose a novel method, the Global Feature Augmented Network (GloFANet), which combines global and local features to guide the model in capturing critical features and improving the discriminability of hard samples. GloFANet is trained using 25,148 images from SoccerNet, along with our own annotated data, totaling 542,690 data points. Experimental results show that GloFANet achieves a mean Average Precision of 85.86% across seven types of keypoints, surpassing the 2023 SoccerNet champion by 5.96%. Notably, in detecting the center line using only 2,137 samples, our method improves performance by 20.48% over the baseline. GloFANet maintains accuracy even with limited data, demonstrating strong data efficiency.