COLMAP Files

New Features

• Added multi-component support to the global mapper. Disconnected view-graph components are reconstructed independently and returned as separate models, including components revealed after filtering outlier relative rotations. The behavior is controlled by GlobalMapper.multiple_models and GlobalMapper.min_model_size and is also available through pycolmap.
• Added LoMa learned feature extraction and matching through ONNX, including LOMA_B and LOMA_B128 descriptors, brute-force matching, and multiple dedicated matcher variants. LoMa is available through the CLI, GUI, and pycolmap, with optional BF16 inference.
• Added ROCm/HIP acceleration for patch_match_stereo, enabling dense reconstruction on supported AMD GPUs through the HIP_ENABLED build option. CUDA and HIP builds are mutually exclusive.
• Added a browser-based, local-only 3D viewer for sparse binary reconstructions, including camera and point inspection, source images, and reprojections.
• Added incremental, global, and hierarchical mapper selection to the GUI, with mapper-specific configuration and progress rendering.
• Added complete pycolmap bindings for hierarchical mapping, including hierarchical_mapping, HierarchicalPipeline, and scene-clustering options. Expanded the global pipeline bindings and callbacks.
• Added 6-point shared-focal and one-sided-focal relative-pose solvers, improving reconstruction when camera intrinsics are unknown or only partially known.
• Added optional DEGENSAC fundamental-matrix estimation for scenes dominated by a plane. It is disabled by default and available through TwoViewGeometry.use_degensac.
• Added nanosecond-resolution timestamps and conversion utilities in C++ and pycolmap.
• Added graceful shutdown for long-running CLI processing pipelines. A first SIGINT or SIGTERM preserves usable intermediate results, while a second signal terminates immediately. PyCOLMAP exposes the same cooperative cancellation through CancellationToken.

Improvements

• Replaced Sampson error on camera bearings with pixel-consistent tangent Sampson error for calibrated two-view geometry, relative-pose refinement, guided matching, and generalized pose estimation. This particularly improves wide-field-of-view and spherical cameras.
• Added camera and image-space point overloads for pycolmap.estimate_relative_pose and refine_relative_pose while retaining the camera-ray overloads for backwards compatibility.
• Estimate homographies on camera rays for distorted and spherical cameras, enabling geometrically correct planar and panoramic detection.
• Refine fundamental matrices with Sampson error by default and use MSAC for robust two-view geometry estimation. Significantly improved accuracy for recovered two-view geometries with small improvements on e2e metrics.
• Added analytical reprojection Jacobians for every camera model and fixed-pose bundle adjustment. Benchmarks show approximately 1.2--1.55x faster incremental mapping for common pinhole and OpenCV camera configurations.
• Added selectable standard or Boost hash-map backends. Recent Boost versions can improve incremental-mapping performance, especially for large scenes.
• Optimized image warping for modest downscales, with reported 4--7x improvements in common undistortion cases. Added configurable nearest-neighbor or bilinear interpolation in C++ and pycolmap.
• Accelerated image preprocessing by using antialiased triangle filtering for bilinear bitmap downscaling and vectorizing RGB-to-grayscale conversion. Bitmap rescaling now also honors the requested bilinear or box filter.
• Added CoreML as an ONNX execution provider on macOS, with automatic CPU fallback for unsupported models.
• Added optional match-count weighting for global rotation averaging.
• Improved sequential pairing for camera rigs by avoiding unrelated cross-sensor temporal pairs.
• Added SequentialMatching.loop_detection_min_index_distance to exclude nearby frames from sequential loop detection without consuming the retrieval budget.
• Added configurable loading of all images during image registration.
• Expanded reconstruction benchmarks with TartanAir panoramas, IMC2025 datasets, multi-seed comparisons, and end-to-end incremental-mapping benchmarks.
• Modernized the documentation site, deployment, landing page, mobile installation selector, and camera-model documentation.

Bug Fixes

• Fix pose-prior alignment and scale preservation for multi-camera rigs.
• Fix spatial matching with missing pose priors and large projected coordinates.
• Fix hierarchical clustering overlap propagation and leaf-size enforcement.
• Fix corrupted PatchMatch source textures when input images have different widths.
• Fix guided matching for spherical cameras.
• Fix Caspar option handling and the fixed-rotation stage of global bundle adjustment.
• Fall back to CPU Ceres bundle adjustment when CUDA support is compiled in but no compatible GPU is available.
• Fix reconstruction merges with inconsistent image ID/name mappings.
• Fix point-triangulator image-list handling.
• Restore the gflags::gflags target for downstream CMake consumers.
• Fix stale reconstruction statistics and elapsed time in the viewer after clearing a reconstruction.
• Fix dense reconstruction from the GUI in HIP-only builds.
• Fix pose-prior bundle adjustment changing rig poses configured as constant.

Breaking Changes

• The global mapper now reconstructs every connected component by default and may therefore return multiple models for disconnected datasets. Set GlobalMapper.multiple_models to false to retain the previous largest-component-only behavior. Models with fewer than GlobalMapper.min_model_size registered frames are discarded.
• The hierarchical_mapper options num_threads, num_workers, image_overlap, and leaf_max_num_images are now prefixed with HierarchicalMapper..
• pycolmap.estimate_essential_matrix no longer returns inlier_points3D.
• pycolmap.cost_functions.SampsonErrorCost now accepts image-plane points instead of camera rays.
• The C++ HierarchicalPipeline::Options type was renamed to HierarchicalPipelineOptions.
• Several C++ relative-pose APIs now use CamRayWithJac.
• PoseFromEssentialMatrix now returns valid correspondence indices instead of triangulated points.
• The C++ CameraModelIsFisheye function was renamed to CameraModelIsPerspectiveFisheye.
• The global std::hash<std::pair<...>> specializations were removed. Downstream unordered containers using pair keys should explicitly use colmap::PairHash.