Segmentation-Based 3D Artistic Rendering

Abstract

This paper introduces segmentation-based 3D non-photorealistic rendering, in which 3D scenes are rendered as a collection of 2D image segments. Segments abstract out unnecessary detail and provide a basis for defining new rendering styles. These segments are computed by a spectral clustering algorithm that incorporates 3D information, including depth, user-defined importance, and object grouping. Temporally coherent animation is created by biasing adjacent frames to have similar segmentations. We describe algorithms for rendering segments in styles inspired by a number of hand-painted images.