Skip to content

Point Cloud Tools v2.3.0 - COLMAP 3.14 (LightGlue + ALIKED)

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Feb 02:52
· 88 commits to master since this release
v2.3.0
288ce88

Point Cloud Tools v2.3.0 - COLMAP 3.14

Pre-built Windows/Linux binaries for COLMAP and pycolmap Python wheels with CUDA 12.8 support.

Important: NVIDIA Driver 570+ Required

CUDA 12.8 binaries require NVIDIA Driver 570 or later. Check with nvidia-smi.

Highlights

  • COLMAP updated to latest — LightGlue ONNX feature matching, ALIKED support, 42 upstream commits
  • GLOMAP removed from repository — Global SfM is now built into COLMAP 3.14 (colmap global_mapper), standalone GLOMAP packages are no longer needed
  • Build system fixes — Disabled ONNX runtime to fix install failures, fixed CMake LIST_SEPARATOR handling
  • Cross-platform — 8 COLMAP packages + 25 pycolmap wheels (Windows & Linux, CPU/CUDA/cuDSS)
  • Python 3.10–3.14 support

COLMAP Upstream Changes (42 commits since v2.2.0)

New Features

  • LightGlue ONNX feature matching (#4098) — State-of-the-art learned feature matching
  • ALIKED support with ONNX (#3593) — Learned local feature detector/descriptor
  • PLY mesh reading (#4065) — Import mesh files directly with roundtrip tests
  • Feature descriptor types (#4080) — Associate descriptors with type and propagate through pipeline
  • Static Windows build with CUDA (#4072)

GLOMAP Integration

  • Moved GLOMAP sources into COLMAP namespace (#4057)
  • Moved GLOMAP Python bindings to colmap folder (#4059)
  • Motion averaging cost functions reorganized (#4058)

Improvements

  • Single pose parameter block and angle axis for BA cost functions (#4029)
  • Abstract general bundle adjuster and solver summary (#4042)
  • Improved Fetzer cost implementation and tests (#4061)
  • Clean up GPS conversion functionality (#4063)
  • Modularized cost functions (#4054)
  • Static Eigen array segment sizes in rotation averaging (#4049)
  • Cleanup outdated pycolmap deprecations (#4081)
  • Fix pycolmap compiler warnings (#4082)
  • Improved test coverage for image reader, rig verification, reconstruction clustering

Build System Changes (this repo)

  • Removed standalone GLOMAP build (PoseLib, colmap-for-glomap no longer needed)
  • Disabled ONNX runtime (-DONNX_ENABLED=OFF) to fix install packaging failures
  • Fixed LIST_SEPARATOR handling for ExternalProject semicolon escaping
  • Updated vcpkg dependencies

What's Included

COLMAP Packages (8 variants)

Package Platform GPU GUI
COLMAP-windows-latest-CPU Windows — —
COLMAP-windows-latest-CUDA Windows CUDA 12.8 —
COLMAP-windows-latest-CUDA-GUI Windows CUDA 12.8 Qt GUI
COLMAP-windows-latest-CUDA-cuDSS Windows CUDA 12.8 + cuDSS —
COLMAP-windows-latest-CUDA-cuDSS-GUI Windows CUDA 12.8 + cuDSS Qt GUI
COLMAP-ubuntu-22.04-CPU Linux — —
COLMAP-ubuntu-22.04-CUDA Linux CUDA 12.8 —
COLMAP-ubuntu-22.04-CUDA-cuDSS Linux CUDA 12.8 + cuDSS —

pycolmap Wheels (25 variants)

Python 3.10–3.14 for Windows and Linux, in CPU / CUDA / CUDA+cuDSS variants.

# Example install
pip install pycolmap-3.14.0.dev0+cuda-cp312-cp312-win_amd64.whl
Variant suffix Description
+cpu CPU-only (no NVIDIA GPU required)
+cuda CUDA 12.8 GPU acceleration
+cuda.cudss CUDA 12.8 + cuDSS sparse solver (fastest bundle adjustment)

Quick Start

COLMAP (Windows):

Expand-Archive COLMAP-windows-latest-CUDA.zip -DestinationPath C:\Tools\COLMAP
C:\Tools\COLMAP\bin\colmap.exe gui

COLMAP (Linux):

unzip COLMAP-ubuntu-22.04-CUDA.zip -d ~/tools/colmap
~/tools/colmap/bin/colmap gui

Global SfM (previously GLOMAP):

colmap global_mapper --database_path db.db --image_path images --output_path sparse

pycolmap:

import pycolmap
pycolmap.extract_features(image_path="images/", database_path="database.db")

Migration from GLOMAP

If you were using the standalone GLOMAP binary from previous releases:

# Old (standalone GLOMAP)
glomap mapper --database_path db.db --image_path images --output_path sparse

# New (COLMAP 3.14+)
colmap global_mapper --database_path db.db --image_path images --output_path sparse

System Requirements

Minimum:

For GPU Acceleration:

  • NVIDIA GPU with Compute Capability 7.5+ (RTX 20/30/40, A100, H100)
  • NVIDIA Driver 570+ required for CUDA 12.8
  • Architectures: 75, 80, 86, 89, 90, 120

For pycolmap:

  • Python 3.10, 3.11, 3.12, 3.13, or 3.14 (64-bit)

Windows Defender: May flag CUDA binaries as false positives (Wacatac.B!ml). Add an exclusion or submit to Microsoft.

License

BSD 3-Clause — COLMAP · Ceres Solver

Built with: Visual Studio 2022, CUDA 12.8, CMake 3.28+, vcpkg

Pre-built COLMAP and pycolmap binaries with CUDA 12.8 support.

Highlights

  • COLMAP updated to latest - Includes LightGlue ONNX feature matching and ALIKED support
  • GLOMAP removed from repository - Global SfM is now built into COLMAP 3.14 (colmap global_mapper)
  • Build fixes - Disabled ONNX runtime to fix install failures, fixed ExternalProject LIST_SEPARATOR handling

Installation

COLMAP

Windows:

Expand-Archive COLMAP-windows-latest-CUDA.zip -DestinationPath C:\Tools\COLMAP
C:\Tools\COLMAP\bin\colmap.exe gui

Linux:

unzip COLMAP-ubuntu-22.04-CUDA.zip -d ~/tools/colmap
~/tools/colmap/bin/colmap gui

pycolmap

Download the appropriate wheel for your Python version and platform, then install:

pip install pycolmap-3.14.0.dev0+cuda-cp312-cp312-win_amd64.whl

Package Variants

Variant Description
CPU CPU-only (no NVIDIA GPU required)
CUDA GPU-accelerated (Windows: self-contained, Linux: requires CUDA Toolkit)
CUDA+cuDSS CUDA + cuDSS sparse solver (faster bundle adjustment)

System Requirements

  • GPU: NVIDIA RTX 20 series or newer (Compute Capability 7.5+)
  • Driver: 570+ (Windows), 570+ (Linux) - CUDA 12.8 requires driver 570+
  • Python: 3.10, 3.11, 3.12, 3.13, or 3.14 (for pycolmap)
Assets 36
Loading