The paper proposes a novel method called Geometry Transfer that utilizes depth maps as style guides to directly stylize the geometry of neural radiance fields, in addition to transferring the appearance. The key highlights are:
Geometry Transfer: The authors introduce the use of depth maps as style guides, in contrast to previous methods that primarily focused on transferring colors and textures. This allows for the direct stylization of the 3D scene's geometry.
Deformation Fields: To ensure coherent stylization of both shape and appearance, the authors employ deformation fields that predict offset vectors for 3D points, enabling synchronized modifications of the geometry and color grids.
RGB-D Stylization: Building on Geometry Transfer, the authors propose novel techniques that leverage both RGB and depth information from the style guide. This includes geometry-aware nearest matching and patch-wise optimization to enhance the diversity and accuracy of the stylization.
Perspective Style Augmentation: The authors introduce a technique to vary the scale of style patterns based on the distance from the camera, enhancing the perception of depth in the stylized 3D scenes.
Partial Stylization: The authors demonstrate the integration of their method with Panoptic Lifting, enabling the selective stylization of target objects or semantic classes within a 3D scene.
Through extensive experiments, the authors show that their proposed Geometry Transfer and RGB-D stylization techniques significantly outperform previous 3D style transfer methods in terms of both qualitative and quantitative evaluations.
Naar een andere taal
vanuit de broninhoud
arxiv.org
Belangrijkste Inzichten Gedestilleerd Uit
by Hyunyoung Ju... om arxiv.org 04-09-2024
https://arxiv.org/pdf/2402.00863.pdfDiepere vragen