Human Pose Estimation in Affective Picture Databases: An Open Dataset of 2D Skeletal Keypoints

Emotionally annotated pictures are typically described primarily through semantic content and by discrete emotion norms or dimensional affective ratings. Structured information about human body posture in such pictures is rarely available in a machine-readable format and must be obtained through additional experiments and data analysis. This limitation is addressed by using human pose estimation on standardized affective image databases and creating an open dataset of two-dimensional skeletal keypoints for images containing visible human figures. Images were selected from the IAPS, NAPS, and OASIS databases based on the presence of people, and OpenPose was used to estimate body keypoints, confidence scores, and the number of detected persons in each image. The resulting dataset contains image identifiers, detected person counts, 2D skeletal keypoint coordinates, and confidence values. In addition to creating the dataset, descriptive statistics and preliminary statistical analyses were performed to explore the relationship between pose-derived features and affective ratings. The main contribution of this paper is an open-source, reusable pose-keypoint dataset that supplements existing affective picture repositories and is intended to support future research on body posture, visual affect, multimodal emotion analysis, and computational models of affective image perception. The created dataset is openly available for research and non-commercial use.