Interacted Planes Reveal 3D Line Mapping.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Robotics

Background:

• 3D line mapping from multi-view RGB images offers a structured scene representation.
• Existing methods often lack explicit physical and topological scene understanding.
• Man-made environments possess inherent geometric structures like planes and lines.

Purpose of the Study:

• To develop an efficient and accurate 3D line mapping framework.
• To integrate physical and topological scene understanding into line mapping.
• To improve structured reconstruction and line-assisted visual localization.

Main Methods:

• LiP-Map: A novel line-plane joint optimization framework.
• Explicitly models learnable line and planar primitives.
• Integrates planar topology via direct primitive interactions, not pairwise constraints.

Main Results:

• Achieves accurate and detailed 3D line mapping with high efficiency (3-5 minutes per scene).
• Demonstrates superior performance on diverse datasets (ScanNetV2, ScanNet++, Hypersim, 7Scenes, Tanks&Temple).
• Significantly advances line-assisted visual localization accuracy.

Conclusions:

• LiP-Map provides a principled approach to structured 3D reconstruction in man-made environments.
• The joint optimization of lines and planes enhances mapping quality and localization.
• The framework offers a strong baseline for future research in geometric scene understanding.