Leveraging Thermal Imaging for Robust Human Pose Estimation in Low-Light Vision

Human Pose Estimation (HPE) is becoming increasingly u-biquitous, finding applications in diverse fields such as surveillance and worker safety, healthcare, sport and entertainment. Despite substantial research in HPE within the visible domain, there is limited focus on thermal imaging for HPE, primarily due to the scarcity and annotation difficulty of thermal data. Thermal imaging offers significant advantages, including better performance in low-light conditions and enhanced privacy, which can lead to greater acceptance of monitoring systems. In this work, we introduce LLVIP-Pose, an extension of the existing LLVIP dataset, to include 2D single-image pose estimation for aligned night-time RGB and thermal images, containing approximately 26k annotated skeletons. We detail our annotation process and propose a novel metric for identifying and correcting poorly annotated skeletons. Furthermore, we present a comprehensive benchmark of top-down, bottom-up, and single-stage pose estimation models evaluated on both RGB and thermal images. Our evaluations demonstrate how pre-training on grayscale COCO data with data augmentation can benefit thermal pose estimation. The LLVIP-Pose dataset addresses the lack of thermal HPE datasets, providing a valuable resource for future research in this area. The pose annotations and baseline code are available on github: https://github.com/MickaelCormier/llvip-pose.