.png){kind=icon}
.png){kind=icon}
Return
Geometry-guided generalizable NeRF for human rendering
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-024-18410-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
It is challenging to render photo-realistic novel views of humans from sparse input views. On one hand, recent works for human rendering are confined to person-specific cases and thus are not generalized to new performers. On the other hand, the algorithms, which are generalizable to novel targets, are developed for scenes or objects and are not directly applicable to novel performers with complex body poses. To this end, we propose a new human rendering pipeline that just takes sparse views of a target performer who never shows up in the training data as the input. Then, it synthesizes high-quality captures at arbitrary viewpoints. The core of our framework is to leverage geometric priors to guide neural radiance fields for human rendering with multi-view images as input. This can not only help deal with the self-occlusion problem caused by skeleton motion when aggregating multi-view features, but also contribute to reasoning about the geometry of the performers. Results of qualitative and quantitative evaluations both show that our method exhibits stronger generalization ability than the current state-of-the-art techniques.
Keywords:
Novel view synthesis
Neural rendering
Image-based rendering
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W