Attention-Based Recurrent Autoencoder for Motion Capture Denoising

Abstract

To resolve the problem of massive loss of MoCap data from optical motion capture, we propose a novel network architecture based on attention mechanism and recurrent network. Its advantage is that the use of encoder-decoder enables automatic human motion manifold learning, capturing the hidden spatial-temporal relationships in motion sequences. In addition, by using the multi-head attention mechanism, it is possible to identify the most relevant corrupted frames with specific position information to recovery the missing markers, which can lead to more accurate motion reconstruction. Simulation experiments demonstrate that the network model we proposed can effectively handle the large-scale missing markers problem with better robustness, smaller errors and more natural recovered motion sequence compared to the reference method.