A Lightweight Optimization Approach to the Single-Person Pose Estimation Pipeline in RGB-D Cameras

The paper presents a systematic benchmark for depth-assisted single-person pose estimation pipelines in three consumer RGB-D cameras.We introduce a lightweight optimization that adjusts only the relative depth coordinates of predicted joints so that their inter-joint depth gaps match those observed in the depth sensor image.The proposed approach is fully differentiable, sensor-agnostic, and light enough for realtime edge deployment, making it immediately applicable to sports coaching, workplace ergonomics, and mixed reality mobile systems.Experiments on a controlled motion capture dataset demonstrate performance trade-offs in accuracy, speed, and robustness under challenging viewing geometries.The findings provide practical guidance on which depth technology best complements state-of-the-art vision models and establish relative depth matching as an effective computationally trivial alternative for laboratory calibration.