COLMAP

COLMAP is a free, open source Structure-from-Motion (SfM) and Multi-View Stereo (MVS) pipeline created and maintained by Johannes Schönberger. Given an unordered set of overlapping photos, it works out where every camera was when each image was taken, producing calibrated camera poses, a sparse 3D point cloud, and optionally dense geometry.

For the radiance field community, COLMAP is the step before training. Most Gaussian Splatting and NeRF workflows depend on it for camera poses and sparse points: trainers such as nerfstudio and Postshot consume its output, the original 3DGS reference implementation expects it, and a growing list of commercial tools either run COLMAP under the hood or read and write its reconstruction format.

What COLMAP actually does

A COLMAP reconstruction moves through a few distinct stages: feature extraction, feature matching, sparse reconstruction with bundle adjustment, and — when you need more than poses — dense Multi-View Stereo, meshing, and texture mapping. Since version 4.0 the front end supports learned ALIKED features through ONNX and LightGlue matching alongside classic SIFT, and it can read EXIF orientation to automatically rotate images during extraction and matching.

The pipeline’s outputs are what downstream tools care about:

• Camera intrinsics and extrinsics — the per-image poses that splat and NeRF trainers align their scenes to

• A sparse point cloud, used by 3DGS training as the initialization for Gaussians

• Optional dense depth and normal maps from PatchMatch Stereo

• Optional meshes, with simplification and texture mapping added in 4.0

The interchange format of splatting pipelines

COLMAP’s sparse reconstruction format has become the lingua franca of the capture ecosystem. RealityCapture added COLMAP export in 1.5, Agisoft Metashape brought COLMAP export to its Standard license, and SplatKing exports native COLMAP data straight from LiDAR captures. Utilities like COLMAP Bridge and the C4D COLMAP converter carry the format into DCC tools. If a product advertises “COLMAP export,” it is advertising compatibility with the entire radiance field training ecosystem.

Incremental vs. global mapping: COLMAP and GLOMAP

COLMAP’s classic incremental mapper grows a reconstruction image by image, registering each new camera against what has already been solved — accurate, but increasingly slow on large captures. GLOMAP took a different approach: a global SfM method built on top of COLMAP that solves all camera poses at once, dramatically faster on big image sets. With COLMAP 4.0, GLOMAP was integrated directly into the main codebase as the global mapper, making global SfM a core reconstruction option rather than a separate companion project.

Recent releases

COLMAP 3.12

Added native sensor-rig support for multi-camera arrays and 360 panorama captures, built-in geo-referencing with latitude/longitude/UTM conversion for drone work, and faster minimal solvers with expanded Python bindings.

COLMAP 4.0

One of the largest updates in the project’s history: GLOMAP’s global SfM pipeline merged in, ALIKED and LightGlue learned matching added, bundle adjustment accelerated, image I/O rebuilt on OpenImageIO, plus model clustering for very large reconstructions, mesh simplification, and texture mapping.

COLMAP 4.1.0

Introduced Caspar, a GPU-accelerated bundle adjustment backend often one to two orders of magnitude faster than the Ceres CUDA solver on medium and large problems, alongside native spherical/equirectangular camera models that reconstruct 360 panoramic captures without rendering perspective views first.

Getting started

COLMAP ships pre-built binaries for Windows, macOS, and Linux with both a GUI and a full command-line interface, and pycolmap exposes the pipeline to Python — the layer where most automated splatting pipelines orchestrate it. The source code lives on GitHub and the documentation covers the full pipeline. It runs CPU-only, though a CUDA GPU substantially accelerates matching, dense stereo, and — since 4.1.0 — bundle adjustment.