Head pose estimation with uncertainty and an application to dyadic interaction detection

Federico Figari Tomenotti, Nicoletta Noceti, Francesca Odone

Abstract: Highlights•We propose a modular pipeline for Head Pose Estimation (yaw-pitch-roll) in RGB images.•We adopt a multi-task heteroscedastic loss obtaining interpretable uncertainty values.•The input of the method are keypoints extracted from 2D human pose estimators only.•Our efficient and light method performs comparably or better than existing approaches.•We apply the pipeline to dyadic interaction detection.