MoRF: Mobile Realistic Fullbody Avatars from a Monocular Video

• DOI: 10.48550/arxiv.2303.10275
• Publication Date: 2023-01-01
• Authors: Alexey Larionov; Evgeniya Ustinova; Mikhail S. Sidorenko; David Svitov; Ilya Zakharkin; Victor Lempitsky; Renat Bashirov
• Keywords: Monocular; Clipping (morphology); Image warping

We present a system to create Mobile Realistic Fullbody (MoRF) avatars. MoRF avatars are rendered in real-time on mobile devices, learned from monocular videos, and have high realism. use SMPL-X as proxy geometry render it with DNR (neural texture image-2-image network). improve prior work, by overfitting per-frame warping fields the neural space, allowing better align training signal between different frames. also refine mesh fitting procedure overall avatar quality. In comparisons other video-based systems, achieve higher image sharpness temporal consistency. Participants of our user study preferred generated MoRF.