CAMERA VIEW-INVARIANT MULTI-PERSON 2D HUMAN POSE ESTIMATION FOR DEPTH AND FISHEYE CAMERAS

Abstract

"This world is gonna be saved by science and art, as well as curious minds and golden hands, and open hearts... Z, 2024

Human pose estimation is the automatic location of the human body parts from an image or video. It is considered a prerequisite for tasks such as activity recognition and human tracking and found its application in human-computer interaction, virtual reality, and sign language recognition. Most of the research on human pose estimation is concentrated on standard RGB cameras, while depth and fisheye cameras were given much less attention due to the lack of corresponding datasets. Since fisheye and depth cameras have several advantages over regular cameras, this project believes that it is important to promote research on human pose estimation for these cameras and proposes to create synthetic camera view-invariant multi-person depth and fisheye image datasets for 2D human pose estimation. Furthermore, it is planned to train state-of-the-art human pose estimation models on these datasets and check their generalization on real images.