On the deployment of Egocentric Pose Estimation Algorithms at the Edge

Egocentric human pose estimation (Ego-HPE) is a key enabler for emerging wearable technologies, where low latency, privacy, and on-device processing are critical requirements. Despite recent advances, most existing approaches are developed and evaluated on high-performance hardware, limiting their practicality in real-world deployments. In this work, we benchmark Ego-HPE algorithms in resource-constrained settings, focusing on execution across heterogeneous platforms ranging from workstation-grade GPUs to embedded edge devices. We compare representative methods across different design paradigms and analyze their trade-offs in accuracy and computational efficiency. The convolutional-based approach achieves an all-keypoint MPJPE of 184.28 mm while maintaining real-time performance on the NVIDIA AGX Orin, with a per-frame latency of 21.93 ms. The results provide practical insights into the design of efficient egocentric HPE systems and outline directions for bridging the gap between high-performance models and deployable on-device solutions.