Model based augmentation and testing of an annotated hand pose dataset

Abstract

Recent advances of deep learning technology enable one to train complex input-output mappings, provided that a high quality training set is available. In this paper, we show how to extend an existing dataset of depth maps of hand annotated with the corresponding 3D hand poses by fitting a 3D hand model to smart glove-based annotations and generating new hand views. We make available our code and the generated data. Based on the present procedure and our previous results, we suggest a pipeline for creating high quality data.