The paper introduces FlexPart, a method for text-controlled, semantics-aware 3D part segmentation. The authors note that a 3D asset can be decomposed in multiple valid ways, and an effective segmentation model should reflect both the desired granularity and the specific parts a user wants to extract.
FlexPart is designed to take text prompts as control signals, allowing users to specify which parts they want. This moves beyond fixed segmentation schemes, offering flexibility in how an object is broken down.
The work addresses a core challenge in 3D understanding: aligning segmentation with semantic intent. By making granularity adjustable, FlexPart aims to be a more practical tool for downstream applications.