Football Players Pose Estimation

The paper presents analysis of algorithms for football players pose estimation based on a custom, real scenario data. Listed approaches have been examined on high resolution videos or photos taken from multiple cameras during football match or training. Wide views has been considered, producing extremely low resolution players' visuals. Tested algorithms consisted of retrained deep convolutional models, optionally retrained for different generalization scenarios. Multiple scenarios for pose estimation compliance have been proposed as well as options for further model augmentations and retraining.