pycolmap-cuda12 3.14.0.dev0

COLMAP bindings

Python Bindings for COLMAP

PyCOLMAP exposes to Python most capabilities of the COLMAP Structure-from-Motion (SfM) and Multi-View Stereo (MVS) pipeline.

Installation

Pre-built wheels for Linux, macOS, and Windows can be installed using pip:

pip install pycolmap

The wheels are automatically built and pushed to PyPI at each release. To benefit from GPU acceleration, wheels built for CUDA 12 (only for Linux - for now) are available under the package pycolmap-cuda12.

[Building PyCOLMAP from source - click to expand]

1. Install COLMAP from source following the official guide.

2. Build PyCOLMAP:

   • On Linux and macOS:

     python -m pip install .
     

   • On Windows, after installing COLMAP via VCPKG, run in powershell:

     python -m pip install . `
         --cmake.define.CMAKE_TOOLCHAIN_FILE="$VCPKG_ROOT/scripts/buildsystems/vcpkg.cmake" `
         --cmake.define.VCPKG_TARGET_TRIPLET="x64-windows"
     

Reconstruction Pipeline

PyCOLMAP provides bindings for multiple steps of the standard reconstruction pipeline:

• Extracting and matching SIFT features
• Importing an image folder into a COLMAP database
• Inferring the camera parameters from the EXIF metadata of an image file
• Running two-view geometric verification of matches on a COLMAP database
• Triangulating points into an existing COLMAP model
• Running incremental reconstruction from a COLMAP database
• Dense reconstruction with multi-view stereo

Sparse & Dense Reconstruction

Sparse & Dense reconstruction from a folder of images can be performed with:

output_path: pathlib.Path
image_dir: pathlib.Path

output_path.mkdir()
mvs_path = output_path / "mvs"
database_path = output_path / "database.db"

pycolmap.extract_features(database_path, image_dir)
pycolmap.match_exhaustive(database_path)
maps = pycolmap.incremental_mapping(database_path, image_dir, output_path)
maps[0].write(output_path)

# Dense reconstruction
pycolmap.undistort_images(mvs_path, output_path, image_dir)
pycolmap.patch_match_stereo(mvs_path)  # requires compilation with CUDA
pycolmap.stereo_fusion(mvs_path / "dense.ply", mvs_path)

PyCOLMAP can leverage the GPU for feature extraction, matching, and multi-view stereo if COLMAP was compiled with CUDA support. Similarly, PyCOLMAP can run Delaunay Triangulation if COLMAP was compiled with CGAL support. This requires to build the package from source and is not available with the PyPI wheels.

Configuration Options

All of the above steps are easily configurable with python dicts which are recursively merged into their respective defaults, for example:

pycolmap.extract_features(
    database_path, image_dir,
    extraction_options={"sift": {"max_num_features": 512}}
)

# Equivalent to:
ops = pycolmap.FeatureExtractionOptions()
ops.sift.max_num_features = 512
pycolmap.extract_features(database_path, image_dir, extraction_options=ops)

To list available options and their default parameters:

help(pycolmap.SiftExtractionOptions)

For another example of usage, see example.py or hloc/reconstruction.py.

Reconstruction Object

We can load and manipulate an existing COLMAP 3D reconstruction:

import pycolmap

reconstruction = pycolmap.Reconstruction("path/to/reconstruction/dir")
print(reconstruction.summary())

for image_id, image in reconstruction.images.items():
    print(image_id, image)

for point3D_id, point3D in reconstruction.points3D.items():
    print(point3D_id, point3D)

for camera_id, camera in reconstruction.cameras.items():
    print(camera_id, camera)

reconstruction.write("path/to/reconstruction/dir/")

Common Operations

The object API mirrors the COLMAP C++ library. The bindings support many operations, for example:

Projecting a 3D point into an image with arbitrary camera model:

uv = camera.img_from_cam(image.cam_from_world * point3D.xyz)

Aligning two 3D reconstructions by their camera poses:

rec2_from_rec1 = pycolmap.align_reconstructions_via_reprojections(
    reconstruction1, reconstruction2
)
reconstruction1.transform(rec2_from_rec1)
print(rec2_from_rec1.scale, rec2_from_rec1.rotation, rec2_from_rec1.translation)

Exporting reconstructions to text, PLY, or other formats:

reconstruction.write_text("path/to/new/reconstruction/dir/")  # text format
reconstruction.export_PLY("rec.ply")  # PLY format

Estimators

We provide robust RANSAC-based estimators for:

• Absolute camera pose (single-camera and multi-camera-rig)
• Essential matrix
• Fundamental matrix
• Homography
• Two-view relative pose for calibrated cameras

All RANSAC and estimation parameters are exposed as objects that behave similarly as Python dataclasses. The RANSAC options are described in colmap/optim/ransac.h and their default values are:

ransac_options = pycolmap.RANSACOptions(
    max_error=4.0,  # For example the reprojection error in pixels
    min_inlier_ratio=0.01,
    confidence=0.9999,
    min_num_trials=1000,
    max_num_trials=100000,
)

Absolute Pose Estimation

To estimate the absolute pose of a query camera given 2D-3D correspondences:

# Parameters:
# - points2D: Nx2 array; pixel coordinates
# - points3D: Nx3 array; world coordinates
# - camera: pycolmap.Camera
# Optional parameters:
# - estimation_options: dict or pycolmap.AbsolutePoseEstimationOptions
# - refinement_options: dict or pycolmap.AbsolutePoseRefinementOptions
answer = pycolmap.estimate_and_refine_absolute_pose(points2D, points3D, camera)
# Returns: dictionary of estimation outputs or None if failure

2D and 3D points are passed as Numpy arrays or lists. The options are defined in estimators/absolute_pose.cc and can be passed as regular (nested) Python dictionaries:

pycolmap.estimate_and_refine_absolute_pose(
    points2D, points3D, camera,
    estimation_options=dict(ransac=dict(max_error=12.0)),
    refinement_options=dict(refine_focal_length=True),
)

Absolute Pose Refinement

# Parameters:
# - cam_from_world: pycolmap.Rigid3d, initial pose
# - points2D: Nx2 array; pixel coordinates
# - points3D: Nx3 array; world coordinates
# - inlier_mask: array of N bool; inlier_mask[i] is true if correspondence i is an inlier
# - camera: pycolmap.Camera
# Optional parameters:
# - refinement_options: dict or pycolmap.AbsolutePoseRefinementOptions
answer = pycolmap.refine_absolute_pose(
    cam_from_world, points2D, points3D, inlier_mask, camera
)
# Returns: dictionary of refinement outputs or None if failure

Essential Matrix Estimation

# Parameters:
# - points1: Nx2 array; 2D pixel coordinates in image 1
# - points2: Nx2 array; 2D pixel coordinates in image 2
# - camera1: pycolmap.Camera of image 1
# - camera2: pycolmap.Camera of image 2
# Optional parameters:
# - options: dict or pycolmap.RANSACOptions (default inlier threshold is 4px)
answer = pycolmap.estimate_essential_matrix(points1, points2, camera1, camera2)
# Returns: dictionary of estimation outputs or None if failure

Fundamental Matrix Estimation

answer = pycolmap.estimate_fundamental_matrix(
    points1,
    points2,
    [options],  # optional dict or pycolmap.RANSACOptions
)

Homography Estimation

answer = pycolmap.estimate_homography_matrix(
    points1,
    points2,
    [options],  # optional dict or pycolmap.RANSACOptions
)

Two-View Geometry Estimation

COLMAP can also estimate a relative pose between two calibrated cameras by estimating both E and H and accounting for the degeneracies of each model.

# Parameters:
# - camera1: pycolmap.Camera of image 1
# - points1: Nx2 array; 2D pixel coordinates in image 1
# - camera2: pycolmap.Camera of image 2
# - points2: Nx2 array; 2D pixel coordinates in image 2
# Optional parameters:
# - matches: Nx2 integer array; correspondences across images
# - options: dict or pycolmap.TwoViewGeometryOptions
answer = pycolmap.estimate_calibrated_two_view_geometry(
    camera1, points1, camera2, points2
)
# Returns: pycolmap.TwoViewGeometry

The TwoViewGeometryOptions control how each model is selected. The output structure contains the geometric model, inlier matches, the relative pose (if options.compute_relative_pose=True), and the type of camera configuration, which is an instance of the enum pycolmap.TwoViewGeometryConfiguration.

Camera Argument

Some estimators expect a COLMAP camera object, which can be created as follows:

camera = pycolmap.Camera(
    model=camera_model_name_or_id,
    width=width,
    height=height,
    params=params,
)

The different camera models and their extra parameters are defined in colmap/src/colmap/sensor/models.h. For example for a pinhole camera:

camera = pycolmap.Camera(
    model='SIMPLE_PINHOLE',
    width=width,
    height=height,
    params=[focal_length, cx, cy],
)

Alternatively, we can also pass a camera dictionary:

camera_dict = {
    'model': COLMAP_CAMERA_MODEL_NAME_OR_ID,
    'width': IMAGE_WIDTH,
    'height': IMAGE_HEIGHT,
    'params': EXTRA_CAMERA_PARAMETERS_LIST
}

SIFT Feature Extraction

import numpy as np
import pycolmap
from PIL import Image, ImageOps

# Input should be grayscale image with range [0, 1].
img = Image.open('image.jpg').convert('RGB')
img = ImageOps.grayscale(img)
img = np.array(img).astype(np.float) / 255.

# Optional parameters:
# - options: dict or pycolmap.SiftExtractionOptions
# - device: default pycolmap.Device.auto uses the GPU if available
sift = pycolmap.Sift()

# Parameters:
# - image: HxW float array
keypoints, descriptors = sift.extract(img)
# Returns:
# - keypoints: Nx4 array; format: x (j), y (i), scale, orientation
# - descriptors: Nx128 array; L2-normalized descriptors

Bitmap

PyCOLMAP provides bindings for the Bitmap class to work with images and convert them to/from NumPy arrays:

import numpy as np
import pycolmap

# Read a bitmap from file
bitmap = pycolmap.Bitmap.read("image.jpg", as_rgb=True)
print(f"Size: {bitmap.width}x{bitmap.height}, Channels: {bitmap.channels}")

# Convert to NumPy array
array = bitmap.to_array()  # Shape: (H, W, 3) for RGB or (H, W) for grayscale

# Create bitmap from NumPy array
array = np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8)
bitmap = pycolmap.Bitmap.from_array(array)

# Write bitmap to file
bitmap.write("output.jpg")

# Rescale bitmap
bitmap.rescale(new_width=320, new_height=240)