Tutorial

3D reconstruction from images traditionally first recovers a sparse representation of the scene and the camera poses of the input images using Structure-from-Motion. This output then serves as the input to Multi-View Stereo to recover a dense representation of the scene.

Quickstart

COLMAP provides an automatic reconstruction tool that simply takes a folder of input images and produces a sparse and dense reconstruction in a workspace folder. Click Reconstruction > Automatic Reconstruction in the GUI and specify the relevant options. The output is written to the workspace folder. For example, if your images are located in path/to/project/images, you could select path/to/project as a workspace folder and after running the automatic reconstruction tool, the folder would look similar to this:

+── images
│   +── image1.jpg
│   +── image2.jpg
│   +── ...
+── sparse
│   +── 0
│   │   +── cameras.bin
│   │   +── images.bin
│   │   +── points3D.bin
│   +── ...
+── dense
│   +── 0
│   │   +── images
│   │   +── sparse
│   │   +── stereo
│   │   +── fused.ply
│   │   +── meshed.ply
│   +── ...
+── database.db

Here, the path/to/project/sparse contains the sparse models for all reconstructed components, while path/to/project/dense contains their corresponding dense models. The dense point cloud fused.ply can be imported in COLMAP using File > Import model from ..., while the dense mesh must be visualized with an external viewer such as Meshlab.

The following sections give general recommendations and describe the reconstruction process in more detail, if you need more control over the reconstruction process/parameters or if you are interested in the underlying technology in COLMAP.

Structure-from-Motion

Incremental Structure-from-Motion pipeline

COLMAP’s incremental Structure-from-Motion pipeline.

Structure-from-Motion (SfM) is the process of reconstructing 3D structure from its projections into a series of images. The input is a set of overlapping images of the same object, taken from different viewpoints. The output is a 3-D reconstruction of the object, and the reconstructed intrinsic and extrinsic camera parameters of all images. Typically, Structure-from-Motion systems divide this process into three stages:

  1. Feature detection and extraction
  2. Feature matching and geometric verification
  3. Structure and motion reconstruction

COLMAP reflects these stages in different modules, that can be combined depending on the application. More information on Structure-from-Motion in general and the algorithms in COLMAP can be found in [schoenberger16sfm] and [schoenberger16mvs].

If you have control over the picture capture process, please follow these guidelines for optimal reconstruction results:

  • Capture images with good texture. Avoid completely texture-less images (e.g., a white wall or empty desk). If the scene does not contain enough texture itself, you could place additional background objects, such as posters, etc.
  • Capture images at similar illumination conditions. Avoid high dynamic range scenes (e.g., pictures against the sun with shadows or pictures through doors/windows). Avoid specularities on shiny surfaces.
  • Capture images with high visual overlap. Make sure that each object is seen in at least 3 images – the more images the better.
  • Capture images from different viewpoints. Do not take images from the same location by only rotating the camera, e.g., make a few steps after each shot. At the same time, try to have enough images from a relatively similar viewpoint. Note that more images is not necessarily better and might lead to a slow reconstruction process. If you use a video as input, consider down-sampling the frame rate.

Multi-View Stereo

Multi-View Stereo (MVS) takes the output of SfM to compute depth and/or normal information for every pixel in an image. Fusion of the depth and normal maps of multiple images in 3D then produces a dense point cloud of the scene. Using the depth and normal information of the fused point cloud, algorithms such as the (screened) Poisson surface reconstruction [kazhdan2013] can then recover the 3D surface geometry of the scene. More information on Multi-View Stereo in general and the algorithms in COLMAP can be found in [schoenberger16mvs].

Preface

COLMAP requires only few steps to do a standard reconstruction for a general user. For more experienced users, the program exposes many different parameters, only some of which are intuitive to a beginner. The program should usually work without the need to modify any parameters. The defaults are chosen as a trade- off between reconstruction robustness/quality and speed. You can set “optimal” options for different reconstruction scenarios by choosing Extras > Set options for ... data. If in doubt what settings to choose, stick to the defaults. The source code contains more documentation about all parameters.

COLMAP is research software and in rare cases it may exit ungracefully if some constraints are not fulfilled. In this case, the program prints a traceback to stdout. To see this traceback or more debug information, it is recommended to run the executables (including the GUI) from the command-line, where you can define various levels of logging verbosity.

Terminology

The term camera is associated with the physical object of a camera using the same zoom-factor and lens. A camera defines the intrinsic projection model in COLMAP. A single camera can take multiple images with the same resolution, intrinsic parameters, and distortion characteristics. The term image is associated with a bitmap file, e.g., a JPEG or PNG file on disk. COLMAP detects keypoints in each image whose appearance is described by numerical descriptors. Pure appearance-based correspondences between keypoints/descriptors are defined by matches, while inlier matches are geometrically verified and used for the reconstruction procedure.

Data Structure

COLMAP assumes that all input images are in one input directory with potentially nested sub-directories. It recursively considers all images stored in this directory, and it supports various different image formats (see FreeImage). Other files are automatically ignored. If high performance is a requirement, then you should separate any files that are not images. Images are identified uniquely by their relative file path. For lather processing, such as image undistortion or dense reconstruction, the relative folder structure should be preserved. COLMAP does not modify the input images or directory and all extracted data is stored in a single, self-contained SQLite database file (see Database Format).

The first step is to start the graphical user interface of COLMAP by running the pre-built binaries (Windows: COLMAP.bat, Mac: COLMAP.app) or by executing ./src/exe/colmap. Next, create a new project by choosing File > New project. In this dialog, you must select where to store the database and the folder that contains the input images. For convenience, you can save the entire project settings to a configuration file by choosing File > Save project. The project configuration stores the absolute path information of the database and image folder in addition to any other parameter settings. If you decide to move the database or image folder, you must change the paths accordingly by creating a new project. Alternatively, the resulting .ini configuration file can be directly modified in a text editor of your choice. To reopen an existing project, you can simply open the configuration file by choosing File > Open project and all parameter settings should be recovered. Note that all COLMAP executables can be started from the command-line by either specifying individual settings as command-line arguments or by providing the path to the project configuration file (see Interface).

An example folder structure could look like this:

/path/to/project/...
+── images
│   +── image1.jpg
│   +── image2.jpg
│   +── ...
│   +── imageN.jpg
+── database.db
+── project.ini

In this example, you would select /path/to/project/images as the image folder path, /path/to/project/database.db as the database file path, and save the project configuration to /path/to/project/project.ini.

Feature Detection and Extraction

In the first step, feature detection/extraction finds sparse feature points in the image and describes their appearance using a numerical descriptor. COLMAP imports images and performs feature detection/extraction in one step in order to only load images from disk once.

Next, choose Processing > Extract features. In this dialog, you must first decide on the employed intrinsic camera model. You can either automatically extract focal length information from the embedded EXIF information or manually specify intrinsic parameters, e.g., as obtained in a lab calibration. If an image has partial EXIF information, COLMAP tries to find the missing camera specifications in a large database of camera models automatically. If all your images were captured by the same physical camera with identical zoom factor, it is recommended to share intrinsics between all images. Note that the program will exit ungracefully if the same camera model is shared among all images but not all images have the same size or EXIF focal length. If you have several groups of images that share the same intrinsic camera parameters, you can easily modify the camera models at a later point as well (see Database Management). If in doubt what to choose in this step, simply stick to the default parameters.

You can either detect and extract new features from the images or import existing features from text files. COLMAP extracts SIFT [lowe04] features either on the GPU or the CPU. The GPU version requires an attached display, while the CPU version is recommended for use on a server. In general, the GPU version is favorable as it has a customized feature detection mode that often produces higher quality features in the case of high contrast images. If you import existing features, every image must have a text file next to it (e.g., /path/to/image1.jpg and /path/to/image1.jpg.txt) in the following format:

NUM_FEATURES 128
X Y SCALE ORIENTATION D_1 D_2 D_3 ... D_128
...
X Y SCALE ORIENTATION D_1 D_2 D_3 ... D_128

where X, Y, SCALE, ORIENTATION are floating point numbers and D_1…D_128 values in the range 0…255. The file should have NUM_FEATURES lines with one line per feature. Note that by convention the upper left corner of an image has coordinate (0, 0) and the center of the upper left most pixel has coordinate (0.5, 0.5). If you must import features for large image collections, it is much more efficient to directly access the database with your favorite scripting language (see Database Format).

If you are done setting all options, choose Extract and wait for the extraction to finish or cancel. If you cancel during the extraction process, the next time you start extracting images for the same project, COLMAP automatically continues where it left off. This also allows you to add images to an existing project/reconstruction. In this case, be sure to verify the camera parameters when using shared intrinsics.

All extracted data will be stored in the database file and can be reviewed/managed in the database management tool (see Database Management) or, for experts, directly modified using SQLite (see Database Format).

Feature Matching and Geometric Verification

In the second step, feature matching and geometric verification finds correspondences between the feature points in different images.

Please, choose Processing > Match features and select one of the provided matching modes, that are intended for different input scenarios:

  • Exhaustive Matching: If the number of images in your dataset is relatively low (up to several hundreds), this matching mode should be fast enough and leads to the best reconstruction results. Here, every image is matched against every other image, while the block size determines how many images are loaded from disk into memory at the same time.

  • Sequential Matching: This mode is useful if the images are acquired in sequential order, e.g., by a video camera. In this case, consecutive frames have visual overlap and there is no need to match all image pairs exhaustively. Instead, consecutively captured images are matched against each other. This matching mode has built-in loop detection based on a vocabulary tree, where every N-th image (loop_detection_period) is matched against its visually most similar images (loop_detection_num_images). Note that image file names must be ordered sequentially (e.g., image0001.jpg, image0002.jpg, etc.). You can verify the correct order in the database management tool (see Database Format). Note that loop detection requires a pre-trained vocabulary tree, that can be downloaded from https://demuc.de/colmap/.

  • Vocabulary Tree Matching: In this matching mode [schoenberger16vote], every image is matched against its visual nearest neighbors using a vocabulary tree with spatial re-ranking. This is the recommended matching mode for large image collections (several thousands). This requires a pre-trained vocabulary tree, that can be downloaded from https://demuc.de/colmap/.

  • Spatial Matching: This matching mode matches every image against its spatial nearest neighbors. Spatial locations can be manually set in the database management. By default, COLMAP also extracts GPS information from EXIF and uses it for spatial nearest neighbor search. If accurate prior location information is available, this is the recommended matching mode.

  • Transitive Matching: This matching mode uses the transitive relations of already existing feature matches to produce a more complete matching graph. If an image A matches to an image B and B matches to C, then this matcher attempts to match A to C directly.

  • Custom Matching: This mode allows to specify individual image pairs for matching or to import individual feature matches. To specify image pairs, you have to provide a text file with one image pair per line:

    image1.jpg image2.jpg
    image1.jpg image3.jpg
    ...
    

    where image1.jpg is the relative path in the image folder. You have two options to import individual feature matches. Either raw feature matches, which are not geometrically verified or already geometrically verified feature matches. In both cases, the expected format is:

    image1.jpg image2.jpg
    0 1
    1 2
    3 4
    <empty-line>
    image1.jpg image3.jpg
    0 1
    1 2
    3 4
    4 5
    <empty-line>
    ...
    

    where image1.jpg is the relative path in the image folder and the pairs of numbers are zero-based feature indices in the respective images. If you must import many matches for large image collections, it is more efficient to directly access the database with a scripting language of your choice.

If you are done setting all options, choose Match and wait for the matching to finish or cancel in between. Note that this step can take a significant amount of time depending on the number of images, the number of features per image, and the chosen matching mode. Expected times for exhaustive matching are from a few minutes for tens of images to a few hours for hundreds of images to days or weeks for thousands of images. If you cancel the matching process or import new images after matching, COLMAP only matches image pairs that have not been matched previously. The overhead of skipping already matched image pairs is low. This also enables to match additional images imported after an initial matching and it enables to combine different matching modes for the same dataset.

All extracted data will be stored in the database file and can be reviewed/managed in the database management tool (see Database Management) or, for experts, directly modified using SQLite (see Database Format).

Note that feature matching requires a GPU and that the display performance of your computer might degrade significantly during the matching process. If your system has multiple CUDA-enabled GPUs, you can select specific GPUs with the gpu_index option.

Sparse Reconstruction

After producing the scene graph in the previous two steps, you can start the incremental reconstruction process by choosing Reconstruction > Start. COLMAP first loads all extracted data from the database into memory and seeds the reconstruction from an initial image pair. Then, the scene is incrementally extended by registering new images and triangulating new points. The results are visualized in “real-time” during this reconstruction process. Refer to the Graphical User Interface section for more details about the available controls. COLMAP attempts to reconstruct multiple models if not all images are registered into the same model. The different models can be selected from the drop-down menu in the toolbar. If the different models have common registered images, you can use the model_converter executable to merge them into a single reconstruction (see FAQ for details). If all your images use the SIMPLE_RADIAL camera model (default) without shared intrinsics, you can use PBA [wu11] instead of Ceres Solver [ceres] for fast bundle adjustment, which can be activated in the reconstruction options under the bundle adjustment section (use_pba=true).

Ideally, the reconstruction works fine and all images are registered. If this is not the case, it is recommended to:

  • Perform additional matching. For best results, use exhaustive matching, enable guided matching, increase the number of nearest neighbors in vocabulary tree matching, or increase the overlap in sequential matching, etc.
  • Manually choose an initial image pair, if COLMAP fails to initialize. Choose Reconstruction > Reconstruction options > Init and set images from the database management tool that have enough matches from different viewpoints.

Importing and Exporting

COLMAP provides several export options for further processing. For full flexibility, it is recommended to export the reconstruction in COLMAP’s data format by choosing File > Export to export the currently viewed model or File > Export all to export all reconstructed models. The model is exported in the selected folder using separate text files for the reconstructed cameras, images, and points. When exporting in COLMAP’s data format, you can re- import the reconstruction for later visualization, image undistortion, or to continue an existing reconstruction from where it left off (e.g., after importing and matching new images). To import a model, choose File > Import and select the export folder path. Alternatively, you can also export the model in various other formats, such as Bundler, VisualSfM [1], PLY, or VRML by choosing File > Export as.... COLMAP can visualize plain PLY point cloud files with RGB information by choosing File > Import From.... Further information about the format of the exported models can be found here.

Dense Reconstruction

After reconstructing a sparse representation of the scene and the camera poses of the input images, MVS can now recover denser scene geometry. COLMAP has an integrated dense reconstruction pipeline to produce depth and normal maps for all registered images, to fuse the depth and normal maps into a dense point cloud with normal information, and to finally estimate a dense surface from the fused point cloud using Poisson reconstruction [kazhdan2013].

To get started, import your sparse 3D model into COLMAP (or select the reconstructed model after finishing the previous sparse reconstruction steps). Then, choose Reconstruction > Multi-view stereo and select an empty or existing workspace folder, which is used for the output and of all dense reconstruction results. The first step is to undistort the images, second to compute the depth and normal maps using stereo, third to fuse the depth and normals maps to a point cloud, followed by a final, optional point cloud meshing step. During the stereo reconstruction process, the display might freeze due to heavy compute load and, if your GPU does not have enough memory, the reconstruction process might ungracefully crash. Please, refer to the FAQ (freeze and memory) for information on how to avoid these problems. Note that the reconstructed normals of the point cloud cannot be directly visualized in COLMAP, but e.g. in Meshlab by enabling Render > Show Normal/Curvature. Similarly, the reconstructed dense surface mesh model must be visualized with external software.

Note that COLMAP requires a CUDA-enabled GPU, so in addition to the internal dense reconstruction functionality, COLMAP exports to several other dense reconstruction libraries, such as CMVS/PMVS [furukawa10] or CMP-MVS [jancosek11]. Please choose Extras > Undistort images and select the appropriate format. The output folders contain the reconstruction and the undistorted images. In addition, the folders contain sample shell scripts to perform the dense reconstruction. To run PMVS2, execute the following commands:

./path/to/pmvs2 /path/to/undistortion/folder/pmvs/ option-all

where /path/to/undistortion/folder is the folder selected in the undistortion dialog. Make sure not to forget the trailing slash in /path/to/undistortion/folder/pmvs/ in the above command-line arguments.

For large datasets, you probably want to first run CMVS to cluster the scene into more manageable parts and then run COLMAP or PMVS2. Please, refer to the sample shell scripts in the undistortion output folder on how to run CMVS in combination with COLMAP or PMVS2. Moreover, there are a number of external libraries that support COLMAP’s output:

Database Management

You can review and manage the imported cameras, images, and feature matches in the database management tool. Choose Processing > Manage database. In the opening dialog, you can see the list of imported images and cameras. You can view the features and matches for each image by clicking Show image and Show matches. Individual entries in the database tables can be modified by double clicking specific cells. Note that any changes to the database are only effective after clicking Save.

To share intrinsic camera parameters between arbitrary groups of images, select a single or multiple images, choose Set camera and set the camera_id, which corresponds to the unique camera_id column in the cameras table. You can also add new cameras with specific parameters. By setting the prior_focal_length flag to 0 or 1, you can give a hint whether the reconstruction algorithm should trust the focal length value. In case of a prior lab calibration, you want to set this value to 1. Without prior knowledge about the focal length, it is recommended to set this value to 1.25 * max(width_in_px, height_in_px).

The database management tool has only limited functionality and, for full control over the data, you must directly modify the SQLite database (see Database Format). By accessing the database directly, you can use COLMAP only for feature extraction and matching or you can import your own features and matches to only use COLMAP’s incremental reconstruction algorithm.

Graphical and Command-line Interface

Most of COLMAP’s features are accessible from both the graphical and the command-line interface. All binaries accept a ./bin -h (help) argument to list the available options. You can provide the options directly as command-line arguments or you can provide a .ini project configuration file containing the options as the ./bin --project_path path/to/project.ini argument. To start the GUI application, please execute ./src/exe/colmap or directly specify a project configuration as ./src/exe/colmap --project_path path/to/project.ini to avoid tedious selection in the GUI. The Graphical User Interface and Command-line Interface sections provide more details about the available controls.

Footnotes

[1]VisualSfM’s [wu13] projection model applies the distortion to the measurements and COLMAP to the projection, hence the exported NVM file is not fully compatible with VisualSfM.