Core Concepts
The author presents S-DyRF as a reference-based spatio-temporal stylization method for dynamic neural radiance fields, aiming to maintain semantic consistency across spatial and temporal dimensions.
Abstract
The content introduces S-DyRF, a novel approach to stylizing dynamic 3D scenes using reference images. The method addresses the challenge of limited stylized references along the temporal axis by generating temporal pseudo-references. It employs coarse and fine style transfer techniques to achieve high-quality results on both synthetic and real-world datasets. The user study and ablation study further validate the effectiveness and superiority of the proposed method.
Stats
"λ = 5 × 10−3"
"λTV = 5 × 10−2"
"λfeat = λc = 1"
"λcolor = 5"
"λf = 10"