Dataset Synthesis for Activity Recognition

Abstract

Novel deep learning models proposed in computer vision are increasingly complex and require increasingly large datasets to be trained. Collecting sufficiently large video datasets is often challenging, both in research and industry. We treat the cases of gait recognition and violence detection, where data collection is often ethically questionable or illegal. This work proposes a powerful and versatile pipeline that can synthesise video data from motion examples to address these and potentially many more issues. Motion examples are either extracted from video using pose estimation techniques or taken directly from motion capture data. We show that using our pipeline 1) synthetic data generation is significantly cheaper than real data collection, 2) synthetic data can be valuable for training, and 3) its value can be measured through comparative experimentation.