In this comparative study, we examine how two different pose estimation pipelines—Orbbec’s proprietary Body Tracking SDK and the open-source OpenPose—affect the quality of reconstructed 3D human body meshes from RGB video streams. Leveraging an Orbbec Astra camera, we capture RGB and depth data to extract keypoints, which are subsequently fed into SMPLify-X. The SMPL and SMPL-X models then generate photorealistic 3D body meshes based on these keypoints. While SMPL focuses exclusively on core body joints, SMPL-X extends its capabilities to include hand pose and facial expression. Our analysis considers the completeness of the resulting meshes, the potential impact of occlusions and rapid movements, and the overall realism of inferred body, hand, and facial poses. By systematically comparing these pipelines and models, we highlight practical limitations—such as partial occlusions and model training data biases—and propose future research directions to enhance accuracy through depth-informed refinements and more robust metrics.

在本对比研究中，我们针对两种不同的姿态估计方案（Orbbec Body Tracking SDK和OpenPose）如何影响基于RGB视频的人体三维重建结果进行了评估。通过Orbbec Astra相机获取的RGB与深度数据，我们提取出关键点并输入SMPLify-X算法，随后使用SMPL与SMPL-X模型生成逼真的人体三维网格。与仅涵盖身体关节的SMPL模型相比，SMPL-X可进一步捕捉手部和面部细节，在关节推断不完整的情况下依然能展现出自然的手部和头部姿态。研究过程中，我们重点关注重建网格的完整性、运动及遮挡对重建质量的影响，以及模型对手部及面部姿态的推断效果。通过系统对比这两套姿态估计管线与三维模型，我们同时也探讨了遮挡场景、数据集偏差以及模型自身局限等现实因素，并提出了在深度信息辅助和评价指标量化方面的未来改进思路。