Peter Kravár
Supervisor(s): Martin Madaras, Lukáš Gajdošech
Masaryk University
Abstract: Methods based on radiance fields have continually pushed the boundaries of novel view synthesis since their first appearance in 2020. The capabilities of radiance fields have since been extended from pure view synthesis to high quality surface reconstruction. However, these surface reconstruction methods generally display poor performance in scenes with transparent objects, producing holes and artifacts in the resulting geometry. Surface reconstruction fails even for radiance fields, which are able to represent the transparent objects accurately in view synthesis. Our work investigates the causes of such failure cases in scenes represented by 3D Gaussians and proposes an improvement to the mesh extraction process. We build upon the method of Gaussian Opacity Fields and utilize it for both scene optimization and geometry extraction. By supplying their geometry extraction pipeline with Gaussians of multiple training steps reached during scene optimization, we achieve a significant uplift in extracted mesh quality for transparent objects. In order to avoid the global reconstruction quality loss incurred by this approach, we segment transparent objects in the scene and selectively apply our pipeline modifications only to those objects.
Keywords: 3D Reconstruction, Geometry Processing, Point-based Graphics, Rendering
Full text: Year: 2025
