Abstract:

Successful object detection, using CNN, requires lots of well-annotated training data which is currently not available for action recognition in hand-ball domain.
Augmenting real-world image dataset with synthesized images is not a novel approach, but the effectiveness of the creation of such a dataset and the quantities of generated images required to improve the detection can be.
Starting with relatively small training dataset, by combining traditional 3D modeling with proceduralism and optimizing generator-annotator pipeline to keep rendering and annotating time under 3 FPS, we achieved 3x better detection results, using YOLO, while only tripling the training dataset.