MeshSeg: 3D Mesh Segmentation Through Color Clustering for Controllable Texture Editing

• Caetano Müller UFRGS
• Gustavo Lopes Tamiosso UFRGS
• Lucas Spagnolo Bombana UFRGS
• Manuel M. Oliveira UFRGS

Resumo

Segmenting 3D meshes is crucial for editing, reuse, and animation of 3D models, but it remains a labor-intensive manual process. We introduce MeshSeg, a 3D segmentation technique that operates on the 2D texture atlas of a 3D model. MeshSeg supports two segmentation strategies: a fast and lightweight classical image-processing pipeline, and a deep learning-based segmentation model (SAM). We encode each texture atlas segment as a color-based vector and merge visually similar regions using unsupervised clustering. An interactive interface further allows users to refine the segmentation through manual adjustments, ensuring high-quality, controllable results. MeshSeg's segmented outputs enable fast, targeted part-based texture editing (e.g., localized recoloring) in downstream applications such as animation, game development, and 3D asset production. By avoiding reliance on training data and focusing solely on color cues, our approach provides a general, easily reproducible solution for textured mesh segmentation across diverse models. We demonstrate its effectiveness on a series of geometrically complex models and show that it frees users from the time-consuming task of manually editing texture atlases.

Palavras-chave: Image segmentation, Solid modeling, Three-dimensional displays, Image color analysis, Pipelines, Training data, Production, Manuals, Animation, Vectors