Synthetic Humans for Action Recognition from Unseen Viewpoints

Abstract Although synthetic training data has been shown to be beneficial for tasks such as human pose estimation, its use RGB action recognition is relatively unexplored. Our goal in this work answer the question whether humans can improve performance of , with a particular focus on generalization unseen viewpoints. We make recent advances monocular 3D body reconstruction from real sequences automatically render videos labels. following contributions: (1) we investigate extent variations and augmentations that are improving at new consider changes shape clothing individuals, well more relevant non-uniform frame sampling, interpolating between motion individuals performing same action; (2) introduce generation methodology, SURREACT allows spatio-temporal CNNs classification; (3) substantially state-of-the-art NTU RGB+D UESTC standard multi-view benchmarks; Finally, (4) extend augmentation approach in-the-wild subset Kinetics dataset case when only one-shot available, demonstrate improvements well.