Self-Supervised Learning of Spatial Object Positioning in Football

We introduce a pretext task for self-supervised learning of feature extraction on an unlabeled dataset of football images. The task is based on predicting the relative distance between two random crops from the same image, which requires the model to understand the spatial positioning of the objects and players in the image. We evaluate the feature extractor trained with the proposed pretext task on the SoccerNet action spotting challenge and compare it to the existing self-supervised method SimCLR. We demonstrate the effectiveness and generality of the proposed pretext task for learning relevant features of the football domain.