μPose: Synthetic Dataset for Human Pose Estimation in Microgravity Environments

The field of space mission has been growing rapidly in recent years, with a multitude of projects and initiatives being undertaken by various organizations around the world. One critical aspect of such endeavors is the ability to simulate and visualize the complex environments and scenarios that space missions may encounter. This paper presents the challenges of human pose estimation in microgravity environments. Existent frameworks face challenges in estimating human poses in non-ordinary positions, and these difficulties are further compounded by the shortage of pertinent data. To overcome these challenges, we propose a synthetic dataset featuring virtual humans in microgravity, serving as a cost-effective alternative for training pose estimation models. Our methodology employs transfer learning with an HRNET backbone and associative embedding, facilitating the retraining of models for 2D human pose estimation on established datasets like COCO and MMPose. Evaluation against publicly available images of astronauts in the International Space Station (ISS) demonstrates our model’s robust generalization, with the best result showing a 20% improvement compared to models trained with COCO dataset. This research demonstrates the viability of synthetic datasets for training models in unique environments. The proposed methodology showcases improved model performance in microgravity scenarios, marking a stride in advancing human pose estimation for space exploration.