ROI estimation and new Viewer

Version 2.4.0 (January 20, 2026)

Summary

• Complete viewer rewrite with OpenGL 3, ImGui UI, multi-select, screenshots, workflows
• Automatic Region-Of-Interest (ROI) detection
• GPU-based selection and improved rendering
• Enhanced dense reconstruction with SfM normals, segmentation from masks
• Support for new formats: VGGSFM, GLTF point clouds
• Python utilities for depth-maps and mesh reconstruction
• Improved CMake and build system

Major Features

Viewer Enhancements

• Multi-select and screenshot functionality - Select multiple objects and capture screenshots
• GPU-based picker implementation - Improved selection performance using GPU
• Console in UI - Integrated console for better debugging and logging
• Workflow support - Run workflows directly from the viewer, including background execution
• Empty scene signaling - Better handling of empty scenes
• File type associations - Improved file type handling on different platforms
• 3D axes drawing - Visual coordinate system display
• OpenGL v3 upgrade with ImGui - Modernized rendering pipeline with new UI framework
• Camera selection - Select and interact with individual cameras
• GLEW replaced with GLAD - Updated OpenGL loader

Dense Reconstruction Improvements

• SfM normals and surface patches estimation - Better initial geometry from sparse reconstruction
• Point-cloud segmentation from image masks - Automatic segmentation support
• GLTF point-cloud loading - Support for GLTF format
• ROI estimation - Automatic Region-Of-Interest detection
• Depth confidence adjustment - Fast neighbor agreement-based confidence tuning
• Point weighting in ROI - Better handling of region of interest
• Depth-map confidence unification - Normalized confidence values between [0,1]

Interface & Import/Export

• VGGSFM output support - Import from VGGSFM reconstruction
• SIMPLE_PINHOLE model for COLMAP - Additional camera model support
• Binary COLMAP format - Export options for binary format, no points, and common intrinsics
• Python scripts:
  • Depth-maps to TSDF mesh
  • Depth-map import from NPY or EXR
  • DMAP save functionality

Build System

• Enhanced CMake - Streamlined dependencies and improved library linking
• vcpkg integration updates - Regular updates to latest vcpkg releases
• Docker to Ubuntu 24.04 - Updated Docker environment

Core Improvements

Mathematics & Transformations

• ConfidenceInterval addition - Statistical analysis support
• SimilarityTransform refactoring - Improved transformation handling
• Scene coordinates normalization - Better numerical stability

Mesh Processing

• Crop to ROI - Mesh cropping within region of interest
• Parameters fine-tuning for thin structures - Better reconstruction of thin objects
• TRasterMesh multithread-safety - Thread-safe mesh operations
• Stricter threshold for degenerate faces - Improved mesh quality
• RemoveDegenerateFaces() and RemoveDuplicatedVertices() - Mesh cleanup utilities

Performance & Stability

• Bug fixes:
  • Tower generation bug fix
  • cList constructor bug (#1229)
  • File class support for files >2GB
• CUDA support:
  • New GPU architectures (A100, etc.)
  • CameraCUDA refactoring
  • Device utilities addition

Add multi-texture support to mesh texturing

Version 2.3.0 (January 7, 2024)

Summary

• store all stages in Interface MVS format by default
• store dense point-cloud as PLY, including all point-views information, etc
• add max-texture-size and multi-texture support to mesh texturing stage
• speedup/improve fix non-manifold mesh algorithm
• unify library/application initialization/destroying
• add python support for reading DMAP and Interface MVS file formats
• account for different pixel center conventions when importing/exporting COLMAP which uses pixel center at (0.5,0.5) instead of (0, 0) as OpenMVS
• add interface to MVSNet, Nerfstudio and RTMV scene formats
• improve interface to Polycam scene format
• speedup PLY load/save
• speedup/improve ray-triangle intersection
• add support for UV coordinates agnostic about orientation
• add support for storing neighbor views information in Interface MVS format
• add support for C++20
• various other fixes

Major Features

Storage & I/O

• Interface MVS format as default - All stages stored in MVS format by default
• Dense point-cloud as PLY - Including all point-views information
• Python API - Read DMAP and Interface MVS file formats from Python
• Speedup PLY load/save - Faster file operations

Texturing

• Max-texture-size support - Control texture resolution
• Multi-texture support - Multiple texture maps for large meshes
• UV coordinates orientation agnostic - Flexible UV handling
• Pixel center convention fixes - Correct pixel center handling

Pipeline Integration

• MVSNet interface - Import/export MVSNet scenes
• Nerfstudio interface - Import Nerfstudio data
• RTMV interface - Support for RTMV format
• Polycam improvements - Better Polycam import

Mesh Operations

• Fix non-manifold speedup/improvement - Faster and better mesh cleanup
• Ray-triangle intersection improvements - More accurate and faster
• Neighbor views in MVS format - Store neighbor information

Build & Development

• C++20 support - Modern C++ features
• Unified initialization - Library/application lifecycle management
• COLMAP pixel center convention - Account for (0.5,0.5) center vs (0,0)

Add Python API

• add Python API
• COLMAP support in MvgMvsPipeline.py
• interface for binary COLMAP
• interface for Polycam scenes
• tower mode #1017
• estimate 3D points scale
• transform scene by a given transform matrix
• unify Docker scripts and add support for GUI
• fix incorrect comparison in BufferedOutputStream #1010
• add lines structure
• compute the focus of attention of a set of cameras
• add image mask support in mesh texturing

Improved densification of textureless areas

• Improve depth-map estimation in textureless areas using multi-resolution
• Add texture sharpening
• Ground plane estimation
• Estimate volume from the reconstructed mesh
• Texture transfer between two meshes
• Automatic Region-Of-Interest estimation
• Simplify build with vcpkg
• Add tests
• Add Continuous Integration action
• Various bug fixes

Various bug-fixes and algorithm fine-tuning

Most important changes:

• fix various bugs and fine-tune Patch-Match CUDA
• update MvgMvsPipeline.py to match the latest OpenMVG and OpenMVS
• add support to rescale a scene, including the images, in InterfaceCOLMAP
• add support to remesh a reconstruction to the mesh cleaning step
• add support to center the scene and set the bounding box to Viewer
• improve view neighbors selection
• generate depth-maps by rendering a given mesh in each view
• simplify view selection in mesh texturing

Faster & more accurate densification

Most important changes:

• new Patch-Match algorithm with improved accuracy and speed (CUDA)
• add geometric-consistency during depth-map estimation
• cluster large scenes to be processed individually (--sub-scene-area)
• Region-Of-Interest (ROI) import and scene trim
• GLTF/GLB exporter

Expose depth-maps

Depth-maps stored as DMAP files can now be exported/imported/visualized. Full changes:

• add DMAP visualization support to Viewer
• expose depth-map file format in Interface.h
• script MvgMvsPipeline.py improvement
• build scripts improvements
• small depth-map estimation improvement
• add option to decimate the input mesh before texturing
• add support for binary COLMAP projects
• add support for Bundle OUT projects (only undistorted cameras)

Implement Semi-Global Matching

Main changes:

• add Semi-Global Matching as a method to estimate a dense point-cloud; this method is still experimental, thus sometimes the speed and completeness might not be as good as the Path-Match approach, though the accuracy could be better; as always any contribution in improving any of these algorithms is welcome

v1.0.1: Merge branch 'develop' into master (#512)

Main changes:

• open depth-map DMAP file format for easy utilization outside OpenMVS (see ExportDepthDataRaw() and ImportDepthDataRaw())
• expose some depth-map functions in SceneDensify.h
• add global ID support to image structure to allow processing large scenes by splitting them into multiple sub-scenes to be processed individually

v1.0

First mature version of OpenMVS.

The quality of the dense point-cloud is much improved. Tested on the two most representative benchmarks in this field, OpenMVS obtains a decent position in the leader-board at only a fraction of run-time compared to the other methods, even though using only a CPU implementation:

• ETH3D Benchmark (created by the COLMAP authors)
• Tanks and Temples Benchmark (created by Intel)

Even better results can be obtained by using also the mesh reconstruction and mesh refinement modules on top of the dense-point cloud and submit a point-cloud sampled from the more accurate mesh.

The support for the OpenMVG SfM solution was extended with an interface to/from the COLMAP solution, supporting both SfM import and dense point-cloud import/export in order for you to experiment with combining different modules at the various stages of the Multi-View Stereo pipeline.

Enjoy!