Kinect-supported dataset creation for human pose estimation

Abstract

Training and evaluation datasets for specific tasks of human pose estimation are hard to find. This paper presents an approach for rapid construction of a precisely annotated training dataset for human pose estimation of a sitting subject, intended especially for aeronautic cockpit. We propose to use Kinect as a tool for collecting ground truth to a purely visual dataset (for reasons defined by the application, use of Kinect or similar structured light-based approaches is impossible). Since Kinect annotation of individual joints might be imprecise at certain moments, manual post-processing of the acquired data is necessary and we propose a scheme for efficient and reliable manual post-annotation. We produced a dataset of 6,322 annotated frames, involving 11 human subjects recorded in various lighting conditions, different clothing, and varying background. Each frame contains one seated person in frontal view with annotation of pose and optical flow data. We used detectors of body parts based on Random Forest on the produced dataset in order to verify its usability. These preliminary results show that the detector can be trained successfully on the developed dataset and that the optical flow contributes to the detection accuracy considerably. The dataset and the intermediary data used during its creation is made publicly available. By this, we intend to support further research and evaluation in the specific topic of human pose estimation focused on a sitting subject in a cockpit scenario.