Improved Hockey Rink Localization via Augmentation and Temporal Frame Analysis
Abstract
Deep-learning based hockey analysis generally requires automaticrink localization from broadcast videos. This information is used to
determine the locations of players and the puck, which is important
for further analysis such as puck trajectory and player behaviour.
Models for this task determine the homography matrix used to warp
the frame onto the rink template, or vice-versa. However, training
models with good performance is challenging due to lack of training
data. Augmentation algorithms have been shown to be effective for
different machine learning tasks. Here we propose a set of new
augmentation techniques specifically for the task of homography
estimation to improve the model’s reliability in new situations. To
further improve smoothness and reliability of localization, we take
advantage of refined homography between successive frames subsampled
from videos in the inference stages. Results show that the
new augmentation technique along with the smoothing approach
can improve the performance by ∼ 2%.