Johannes Schönberger has released COLMAP 4.1.0, the major structure from motion pipeline that most Gaussian splatting and NeRF workflows depend on to produce camera poses and sparse point clouds. With this new version comes Caspar, a GPU accelerated bundle adjustment backend that is often one to two orders of magnitude faster than the Ceres CUDA solver for medium to large scale problems. Bundle adjustment is typically the slowest stage on large captures, and because the incremental mapper invokes it repeatedly as it grows a reconstruction, the speedups Caspar delivers compound into drastically faster mapping rather than a one time win at the end.
Caspar is selectable as an alternative to the default Ceres solver rather than a replacement, so existing pipelines keep their behavior unless they opt in, and it ships with rig support, GPU device selection, and pycolmap bindings for choosing both the bundle adjustment and the mapper backends from Python. The layer where most automated splatting pipelines actually orchestrate COLMAP. The prior 4.0 release laid the infrastructural groundwork for this kind of modularity, with the global mapper and broader pipeline reorganization. 4.1.0 builds on that by making the solver itself a swappable component. Where Ceres on CUDA was the fast path before, it was still the bottleneck on scenes.
The second structural change is native support for spherical, equirectangular camera models, which enables reconstruction of 360 panoramic images without first rendering them into perspective views. The equirectangular path is generally faster but less accurate than rendering perspective views, as performed in the panorama_sfm example and the panorama_sfm example was itself extended to convert perspective cameras back to equirectangular and to support the global mapper. This removes the perspective cropping and re-stitching step that previously stood between a panoramic capture and a COLMAP reconstruction, a workflow that has grown more common as 360 capture feeds into splatting alongside conventional photogrammetry exports like those in RealityCapture’s COLMAP export path. Rounding out the camera-model work, 4.1.0 adds the Enhanced Unified Camera Model (EUCM), introduces a CamRayFromImg bearing-vector unprojection interface with bindings, and teaches the p4pf solver to estimate separate fx and fy for two focal camera models.
Beyond the solver and camera models, the release adds advancing front surface reconstruction meshing, support for extracting gravity pose priors from EXIF orientation tags, and a set of mapper controls including GlobalMapper.keep_max_num_tracks and rotation averaging options in the global mapper, alongside a version CLI command, an MVS_ENABLED build option, pycolmap bindings for ReprojectionErrorType, a camera up vector in the viewer, Material Symbols GUI icons, and import/export keyboard shortcuts. The matching and triangulation paths were tuned as well. The exhaustive matcher was accelerated by switching from IndexIVFFlat to IndexIVFScalarQuantizer, color extraction was parallelized, and the incremental triangulator was made faster through reused BFS allocations and a simplified merge_trials_, while gravity-aligned rotation averaging was strengthened against 180-degree flips.
4.1.0 fixes empty PatchMatch results on Blackwell GPUs (sm_100+), which had broken multi view stereo on the newest NVIDIA cards, and it repairs the broken integration of SIFT features with the LightGlue matcher, a crash in rotation averaging with multi-image rigs, and Windows Unicode path handling in Bitmap::Read and Bitmap::Write.
All of these updates should make it both materially faster on large scenes and capable of ingesting 360 captures natively, with the Blackwell fix restoring dense reconstruction on the latest GPUs.
It is available now through GitHub.
