COLMAP

An open-source 3D reconstruction pipeline combining structure-from-motion and multi-view stereo.

COLMAP is a general-purpose 3D reconstruction pipeline created by Johannes Schönberger and others. It first runs structure-from-motion (SfM — recovering camera poses and a sparse point cloud simultaneously from multiple photos), then multi-view stereo (MVS — densifying that into a full point cloud and mesh). It matters because it has become close to the de facto standard for “getting camera poses from photos”: training data for neural radiance fields and 3D Gaussian splatting is almost always pose-labeled using COLMAP first. It ships with both a command-line and a graphical interface, and is also commonly called from scripts; in robotics, it's often used to turn a video taken circling an object or scene into a usable 3D asset.

ExampleBefore training a 3D Gaussian splatting model, run a set of images taken circling a tabletop through COLMAP to get camera intrinsics, extrinsics, and a sparse point cloud.

Related

Structure from Motion · Multi-View Stereo · 3D Gaussian Splatting · Neural Radiance Fields · Bundle Adjustment · Camera Extrinsics

Sources

COLMAP Documentation
colmap/colmap - GitHub