Conceitos essenciais
A novel geometry-based keypoints grid and a robust pipeline for 3D camera calibration and homography estimation in sports broadcast videos.
Resumo
The paper proposes a novel framework for 3D sports field registration, particularly in the soccer domain. The key contributions are:
A geometry-based keypoints grid and a robust pipeline for their retrieval, leveraging the known dimensions and markings of the sports field.
A calibration pipeline capable of integrating non-planar points (e.g., goal posts, crossbars) for 3D camera calibration and extending to multiple views from the broadcast.
A minimalist approach focused solely on 2D-3D correspondences, without further refinement.
The proposed method is evaluated on three real-world soccer broadcast datasets (SoccerNet-Calibration, WorldCup 2014, and TS-WorldCup). It demonstrates superior performance in 3D camera calibration compared to state-of-the-art methods, while also achieving competitive results in homography estimation.
The authors first model the soccer field and define a hierarchical structure to compute a pre-defined set of keypoints based on the field's geometric properties. These keypoints are then detected using encoder-decoder convolutional neural networks. The estimated keypoints are used to compute the projection matrix using the Direct Linear Transformation (DLT) algorithm and RANSAC.
The paper also addresses challenges such as keypoint disambiguation, left-right field disambiguation, and handling of non-planar points for robust 3D camera calibration. The authors conduct extensive experiments and provide detailed quantitative and qualitative results, demonstrating the effectiveness of their approach.
Estatísticas
The paper reports the following key statistics:
Accuracy@5 (Acc@5) of 75.3% on the SN22-test-center dataset for camera calibration.
Median reprojection error of 0.011 on the WorldCup 2014 test dataset for homography estimation.
Median projection error of 0.20 meters on the TS-WorldCup test dataset for homography estimation.
Citações
"A novel geometry-based keypoints grid and a robust pipeline for their retrieval."
"A calibration pipeline capable of integrating non-planar points for 3D camera calibration and extending to multiple views from the broadcast."
"A minimalist approach focused solely on 2D-3D correspondences, without further refinement."