Deep motion refinement for stitched locomotion

Abstract

Although motion capture technologies or manual keyframing guarantee high quality results, every possible motion data cannot be created with those methods because it is highly time consuming. Therefore, a few crucial animations are created and then connected with motion blending algorithms to produce motion sequences needed for the final product. However, conventional motion blending algorithms based on interpolation struggle to produce natural locomotion in some cases. To overcome this limitation, we propose a data-driven motion refinement framework based on recurrent neural networks(RNN). Our framework takes a naively stitched locomotion sequence as input, and refines it to a natural locomotion sequence. We also designed a novel data pair generation based on nearest neighbor search. After training, our model produces higher quality results compared to a conventional motion blending algorithm by adjusting the steps rather than simply interpolating positions or rotations. The trained model is also capable of refining various lengths of motion sequences, which allows it to be applied to many applications that require natural motion stitching.