Add to Chrome
Add to Firefox
Add to EdgeInteractive Segmentation of Radiance Fields
Dec 27, 2022
Radiance Fields (RF) are popular to represent casually-captured scenes for new view generation and have been used for applications beyond it. Understanding and manipulating scenes represented as RFs have to naturally follow to facilitate mixed reality on personal spaces. Semantic segmentation of objects in the 3D scene is an important step for that. Prior segmentation efforts using feature distillation show promise but don't scale to complex objects with diverse appearance. We present a framework to interactively segment objects with fine structure. Nearest neighbor feature matching identifies high-confidence regions of the objects using distilled features. Bilateral filtering in a joint spatio-semantic space grows the region to recover accurate segmentation. We show state-of-the-art results of segmenting objects from RFs and compositing them to another scene, changing appearance, etc., moving closer to rich scene manipulation and understanding. Project Page: https://rahul-goel.github.io/isrf/
