CAD-Prompted SAM3: Geometry-Conditioned Instance Segmentation for Industrial Objects

Zhenran Tang, Rohan Nagabhirava, Changliu Liu · 2026

Read CAD-Prompted SAM3: Geometry-Conditioned Instance Segmentation for Industrial Objects by Zhenran Tang, Rohan Nagabhirava, Changliu Liu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-20551.txt.

Description

Word-prompted segmentation hits a wall with objects that words cannot pin down, a common case in factories and 3D printing shops. Zhenran Tang and two co-authors note that image examples also fall short, since they encode looks like color and texture rather than a part's shape. One part may come in many materials, finishes, or colors, so look-based prompts mislead. Yet such parts are usually defined by exact CAD models that hold their true shape. The paper builds a CAD-prompted method on SAM3.

The method feeds multi-view renders of a CAD model as prompts, giving shape-based guidance that ignores surface looks. Training uses synthetic data made from mesh renders in simulation across varied views and scene settings. The result is single-stage, CAD-prompted mask prediction, stretching promptable segmentation to objects that neither words nor looks alone can robustly describe.

Catalog record

CREATORZhenran Tang, Rohan Nagabhirava, Changliu Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-20551

Browse the full library · About the Tower of Babel