POSITION: Open World 3D Scene CAD Recomposition.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computational Geometry

Background:

• 3D scene CAD recomposition reconstructs environments using CAD models.
• Current methods struggle with generalization due to limited training data.
• Reconstructing diverse real-world scenes remains a challenge.

Purpose of the Study:

• To develop an open-world 3D scene CAD recomposition method.
• To improve generalization to diverse scenes and scalable CAD databases.
• To accurately simulate geometric properties and spatial arrangements of original environments.

Main Methods:

• POSITION employs a divide-and-conquer strategy.
• Extracts open-world multi-modal object representations from 3D scenes.
• Utilizes coarse-to-fine retrieval for visually, geometrically, and semantically matching CADs.
• Applies physically plausible pose alignment for consistent geometry and layout.

Main Results:

• Achieves generalization across various scene types and scalable CAD databases without retraining.
• Demonstrates superior CAD recomposition performance on Scan2CAD and real-world datasets.
• Successfully reconstructs 3D scenes with CADs from an open-set database.

Conclusions:

• POSITION offers a robust solution for open-world 3D scene CAD recomposition.
• The method overcomes limitations of existing scan-to-CAD approaches.
• Enables accurate simulation of geometric properties and spatial arrangements in diverse environments.