Human Pose Estimation Using Thermal Images
Javier Smith, Patricio Loncomilla, Javier Ruíz-del-Solar
Abstract
This study addresses the human pose estimation problem on thermal images using Convolutional Neural Networks and Vision Transformer architectures. To do this, eight human pose estimation methods designed for visible images were extended to be applied in the thermal domain. Due to the lack of large, representative datasets containing labeled thermal images, this extension requires transfer learning between the visible and the thermal domain, and a database for fine-tuning the networks in the thermal domain. Thus, it is proposed to train the networks using a grayscale version of the COCO dataset, and then fine-tune them in the thermal domain. Fine-tuning is carried out using the new UCH-Thermal-Pose database presented in this work. This database includes 600 thermal images for training, 200 for validation, and 104 for testing, all of them fully labeled. Moreover, in the paper, a comparative study of the eight extended deep-based methods for human pose detection is carried out. The UCH-Thermal-Pose database and the source code of all the methods compared are made available for future studies.