Variational Pose Prediction with Dynamic Sample Selection from Sparse Tracking Signals

We propose a learning-based approach for full-body pose reconstruction from extremely sparse upper body tracking data, obtained virtual reality (VR) device. leverage conditional variational autoencoder with gated recurrent units to synthesize plausible and temporally coherent motions 4-point (head, hands, waist positions orientations). To avoid synthesizing implausible poses, we novel sample selection interpolation strategy along an anomaly detection algorithm. Specifically, monitor the quality of our generated poses using algorithm smoothly transition better samples when falls below statistically defined threshold. Moreover, demonstrate that method can be used other applications, such as target hitting collision avoidance, where should adhere constraints environment. Our system is lightweight, operates in real-time, able produce realistic motions.