Efficient and Effective Dynamic Hair Reconstruction via Strand Gaussians

Wentao Yang, Chentao Shen, Zejian Kang, Xiangru Huang

Keywords: Dynamic Hair Reconstruction, Efficient Hair Reconstruction

Abstract: Reconstructing dynamic human hair is a crucial way to acquire high quality 3D assets that empowers downstream tasks such as human avatars and animation. Recently, several research works propose using strand gaussians to model human hair, leading to superior reconstruction quality. However, their optimization algorithms admit significant computational burden and typically take hours to days reconstruct dynamic human hair from multi-view videos. We propose an efficient and effective optimization method that is significantly faster compared to state-of-the-art dynamic human hair reconstruction methods, while achieving comparable reconstruction quality. The performance of our algorithm is demonstrated both qualitatively and quantitatively on the public NeRSemble dataset.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 12313