Deep learning based ball detector

Abstract: Sports analysis that traditional computer vision techniques have long dominated is today getting replaced with more advanced machine learning models. To provide analysis in sports, tracking methods have to be fast and reliable, both for ease of use and for broadcasting systems to deliver data to customers quickly. With the growth of large datasets and the rapid development of Graphics Processing Units (GPUs), machine learning models are getting better and more precise. In football games, efficiently tracking the ball is essential when gathering statistics and performing event detection such as offside or passes. This thesis explores football detection with high-resolution images by extending current Deep- Ball and High-Resolution Net in three ways, using Gaussian labels to tackle inaccurate annotations, encoding temporal information with multi-frame input, and providing context by training on player segmentation masks. The results show that using Gaussian labels can help improve the performance drastically in some cases, especially when assuming that there only exists one ball per image. Using multiple frames proved to detect balls in challenging images, where the corresponding single-frame models failed. After training on player segmentation masks, the models were able to detect players accurately but did not show any significant improvement in terms of ball detection. Surprisingly when comparing DeepBall and High-Resolution Net, the results showed that even though DeepBall consist of far less parameters, it achieves higher performance in many of the cases.