Occlusion-aware heatmap generation for enhancing 3D human pose estimation in multi-person environments

In multi-person 3D human pose estimation (HPE), the lack of diverse and accurate 3D pose datasets remains a critical challenge. Despite recent advancements in learning-based methods, real-world scenarios with varied environments and individuals often lead to data biases and sparse annotations, complicating the achievement of robust generalization in visual computing applications. While recent data augmentation methods have shown promise in enhancing the generalization of 3D HPE, the majority target single-person settings, leaving multi-person scenarios insufficiently covered. Our paper presents a novel data augmentation technique for multi-person 3D HPE. We refine the data evolution framework to generate new single-person 3D poses and then combine them into multi-person scenarios. Notably, our method generates occlusion-aware 2D heatmaps by considering camera positions, 3D poses, and joint-specific occlusion uncertainties, capturing the nuances of real-world pose challenges. Evaluations on well-known datasets, such as CMU Panoptic, Shelf, and Campus, demonstrate our method’s effectiveness, especially in constrained data environments. The code and dataset are available at: https://github.com/hyeon0819/MPDA.