Video-based sports activity recognition for children

Large-scale action recognition datasets contain more instances of adults than children, and models trained with these datasets may not perform well for children. In this study, we test if current state-of-the-art deep learning models have some systemic bias in decoding the activity being performed by an adult or a child. We collected a sports activity recognition dataset with child and adult labels. We fine-tuned a state-of-the-art action recognition classifier on two different segments of our dataset, containing only children or only adults. Our results show that cross-condition generalization performance of the resulting networks is not similar. Our results indicate that the child-specific segment is more complex to generalize than the adult-specific segment. The dataset and the code are made publicly available.