Viewpoint-Invariant Soccer Pitch Registration Using Geometric and Learned Features

Description

This website provides supplementary material associated with the registration strategy proposed in [1].

The proposed method performs fully automatic registration of broadcast soccer images to a standardized field model. It combines three complementary geometric cues—white field markings (lines and elliptical arcs), grass-band delimitation lines, and the playing-field mask—and applies a structured labeling process to drastically reduce correspondence ambiguity. By exploiting intersection patterns and projective invariants, the system generates candidate homographies, evaluates them quantitatively through reprojection accuracy and mask consistency, and selects the optimal solution.

This viewpoint-invariant strategy achieves highly accurate and robust registration results across challenging scenarios, including diverse stadiums, multiple zoom levels, varied lighting conditions, and extreme camera perspectives.

Citation

[1] C. Cuevas, D. Berjón, and N. García, “Viewpoint-Invariant Soccer Pitch Registration Using Geometric and Learned Features", Journal of Visual Communication and Image Representation, under review.