Použití self-supervised learning pro rozpoznání sportovních pozic v obraze

This thesis demonstrates a solution for minimizing the amount of necessary labelled training data in the classification of sports poses using a neural network trained with contrastive self-supervised learning. Training consists of two stages. The first stage trains a feature extractor which uses unlabelled training images extracted from recordings of exercises from multiple viewpoints. In the second stage, using a small amount of labelled data, a simple classifier connected to the feature extractor is trained. The thesis discusses classification in the context of yoga poses, however, the final solution can be easily applied to any other sport in case of obtaining a suitable dataset. During the development of the solution, emphasis is placed on the performance of the resulting model so that it can be used on mobile devices. The resulting model reached an accuracy of 76 % using augmentations with a data set containing four labelled images per yoga pose. On a larger data set with 800 labelled images for all poses, an accuracy of 82 % is reached. 

image recognition, pose estimation, contrastive learning, self-supervised learning