Repository navigation
LIMAP 2.0.0 turns this open-source project from a line mapping library into a holistic 3D visual mapping and localization toolbox. Notably, the release brings in the incremental SfM pipeline introduced in Robust Incremental Structure-from-Motion with Hybrid Features (ECCV 2024) and further improved in Stable and Scalable Bundle Adjustment of Holistic 3D Structures (ECCV 2026), and unifies the whole toolbox with the COLMAP ecosystem.
Unification with the COLMAP ecosystem
The point side of the pipeline is COLMAP now, including the scene types, camera models, database, correspondence graph, various robust estimators and the incremental mapper logic. LIMAP adds structures on top instead of maintaining a parallel implementation. The unification is the result of substantial engineering efforts on both ends.
- A reconstruction is written as a plain COLMAP model —
cameras.bin,images.bin,points3D.bin,rigs.bin,frames.bin— with the line, group and wireframe structures alongside it understructures/. There is no conversion step and no LIMAP-specific container in between, so any output can be opened directly in the COLMAP GUI and read back withpycolmap. We also provide additional data structures for reading the new holistic reconstruction format, built by composition around the COLMAP reconstruction object. - Line, group and point correspondences ride in the same
TwoViewGeometrystructures, split over COLMAP'sdatabase.dband LIMAP's
structure_database.db. - Active advances on the COLMAP side carry over directly by bumping the pinned
FetchContenttag: multi-camera rig support, improved two-view geometry estimation, newly introduced utilities, and upstream bug fixes. As of the release date, LIMAP pins COLMAP 4.2.0, released in the same week.
Holistic incremental SfM
- New holistic incremental mapper that reconstructs from images alone, jointly optimizing points, lines, vanishing points, planes and the wireframe together with the camera poses (
limap.cli.structure_incremental_sfm, orautomatic_structure_incremental_reconstructionfor frontend + mapper). - Uncalibrated input supported end to end: focal lengths are estimated jointly with the poses, in both SfM and localization. The underlying PnPL and PnPLF solvers were thoroughly tested and upstreamed to PoseLib, and are used directly through wrapped interfaces.
- Consistent accuracy gains over COLMAP given the same sparse point features and matches, particularly in structured indoor environments. The runtime overhead over COLMAP is ~30% on large scenes, with similar scaling behavior.
HolisticReconstructionenables unified I/O across triangulation, localization and SfM. With poses given,limap.cli.automatic_structure_triangulationautomatically reconstructs vanishing points, planes and the wireframe on top of points and lines.
Structures beyond lines
- Groups as a first-class abstraction — vanishing points, planes and parametric primitives (spheres, cylinders, ellipsoids, cuboids, cones), with 3D estimators and verification for each.
- The wireframe connecting the structures is reconstructed and optimized alongside them.
- New stable and scalable holistic bundle adjustment over all the 3D structures (ECCV 2026), replacing the previous line-only refinement.
- Integrated the new plane detector PxwPlanar, from Pixel-wise Planarity for High-Precision Monocular Plane Segmentation (ECCV 2026).
Packaging and platforms
- Published on PyPI as
pylimap—pip install pylimapgets the compiled core with no build. The import name stayslimap. - Wheels for Linux, macOS and Windows, Python 3.10–3.13. All three platforms are tested in CI.
- Third-party dependencies (COLMAP, PoseLib, JLinkage, libigl) are all pulled with
FetchContentat pinned commits; git submodules are gone.
Codebase and API
- The configurations and implementations of many modules were modernized, improving extensibility and making the toolbox easier to maintain.
- A dedicated
limap.geometrypackage abstracts the line-related geometric operations. - All the pipelines are entered through the new thin
limap.clipackage withpython -m limap.cli.<name>, which only resolves arguments and YAML configs into the option dataclasses. The orchestration stays inlimap.runnersand the computation in C++. limap.estimatorsnow collects all the estimation code behind a common interface: absolute pose, triangulation, 3D lines, 3D groups and the holistic bundle adjustment.- The pybind11 bindings are fully separated from the implementation.
src/limap/<module>/holds the C++ core together with the Python code of the same module, whilesrc/pylimap/<module>/holds nothing but its bindings, mirroring the same layout. - The design follows COLMAP's principles throughout, leaving the door open for further consolidation.
See the documentation for the tutorials and the full API.