cvg/limap

What it solves

LIMAP addresses the limitations of traditional 3D reconstruction (Structure from Motion), which typically relies only on point clouds. By treating lines, vanishing points, planes, and parametric primitives (like spheres and cylinders) as first-class geometric entities, it creates more holistic and accurate 3D models and camera pose estimations than point-only methods.

How it works

LIMAP jointly optimizes camera poses and structural constraints using a variety of geometric features. It integrates several specialized tools for line detection, matching, and plane estimation. The system is fully compatible with the COLMAP ecosystem, utilizing COLMAP's point-based pipeline while adding structural layers (lines, groups, and wireframes) on top of the standard model.

Who it’s for

It is designed for computer vision researchers and engineers working on 3D mapping, visual localization, and structure from motion (SfM), particularly those needing high-precision reconstructions of environments containing structured geometric primitives.

Highlights

• Three Core Pipelines: Supports visual mapping/triangulation, visual localization (including uncalibrated queries), and holistic incremental SfM.
• COLMAP Integration: Fully compatible with COLMAP 4.2.0, allowing reconstructions to be opened in the COLMAP GUI and read via pycolmap.
• Extensible Architecture: Uses registries for line detectors, matchers, vanishing point estimators, and plane detectors to easily incorporate new algorithms.
• Hybrid Feature Support: Jointly optimizes points and lines to improve pose accuracy and stability, especially in uncalibrated scenarios.