Depth-Only Human Pose Estimation

초록

Human pose estimation typically relies on RGB data, which raises privacy concerns due to sensitive information such as facial identities. In addition, RGB-based approaches are highly sensitive to environmental factors, including lighting conditions and complex backgrounds. To address these limitations, this paper proposes a depth-only top-down 2D human pose estimation method. Specifically, we fine-tune RT-DETRv2 for human detection and HRNet for pose estimation to be compatible with depth data. Considering the scarcity of high-quality, universally applicable depth datasets, we converted the RGB-based MS COCO dataset into depth images using Depth Anything v2. Furthermore, to enhance the accuracy of the detection model, pseudo labels were generated from Infrared Radiation images precisely aligned with the depth sensor, utilizing OpenSeeD and MaskDINO models. Our approach significantly improves the performance of depth-based human pose estimation and offers solutions to the shortage of high-quality depth datasets. © 2025 IEEE.

키워드