中性運動解纏於異質性人體運動風格轉換

Abstract

人體運動風格轉換是遊戲設計師和動畫師能應用的重要技術，可以增強運動片段中的角色表現力。然而，目前的數據驅動方法要麼需要配對的運動數據，要麼會由於模型對於運動的解纏不良造成產生的結果不具有所需的內容和風格。因此本研究提出了一種新穎的中性運動解纏(NMD)模型，可以產生高質量的人體運動風格轉換結果，特別是在異質性運動之間。我們的框架通過對抗式的訓練過程，同時學習重建中性運動和輸入運動來分解輸入運動的內容和風格。此外，我們提出了一種新穎的分群損失，進一步增強了內容和風格潛在空間的解纏。我們通過全面的實驗、消融研究和用戶研究評估了我們的方法。結果表明，特別是在異質性運動之間轉換風格時，我們的方法可以生成比現有的運動風格轉換更好的結果。

Motion style transfer is a valuable technique for game designers and animators that enhances the expressiveness of animation clips. However, current data-driven methods either require paired motion data or produce results that do not have desired content and style, due to poor motion disentanglement. This work presents a novel neutral motion disentanglement (NMD) model that produces high-quality motion style transferred results, especially between heterogeneous motions. Our framework decomposes the content and style of an input motion by learning to reconstruct the neutral motion and the input motion simultaneously through an adversarial training process. Additionally, we propose a novel clustering loss that further enhances the disentanglement of the content and style latent spaces. We evaluate our method through thorough experiments, ablation studies, and user studies. The results suggest that our method can generate better transferred results than existing motion style transfer, particularly for transferring styles between heterogeneous motions.