Bag of Tricks Toward Enhanced Shuttlecock Detection from Badminton Videos

Shuttlecock detection is crucial for video analysis in badminton matches, but identifying tiny and fast-moving shuttlecock from videos is still challenging. In this work, we propose several tricks that can be attached to a state-of-the-art shuttlecock detection algorithm, TrackNetV2. Our proposed tricks include introducing residual network, Leaky ReLU, and post-addition batch normalization, all based on the deep learning theory. The contribution of this work is to apply the well-known techniques to shuttlecock detection successfully, and this is the first attempt in shuttlecock detection literature. Our experiments on a public badminton dataset demonstrate that our proposed tricks can substantially improve TrackNetV2 regarding precision, recall, F1, and accuracy from 84.8%, 81.8 % ,83.3 % , and 75.8% to 88.2%, 83.1%,85.3 % , and 78.1%, respectively.