Binary File Format¶
Note that all binary data is stored using little endian byte ordering. All x86 processors are little endian and thus no special care has to be taken when reading COLMAP binary data on most platforms.
Indices and Identifiers¶
Any variable name ending with
*_idx should be considered as an ordered,
contiguous zero-based index. In general, any variable name ending with
should be considered as an unordered, non-contiguous identifier.
For example, the unique identifiers of cameras (CAMERA_ID), images (IMAGE_ID), and 3D points (POINT3D_ID) are unordered and are most likely not contiguous. This also means that the maximum POINT3D_ID does not necessarily correspond to the number 3D points, since some POINT3D_ID’s are missing due to filtering during the reconstruction, etc.
By default, COLMAP uses a binary file format (machine-readable, fast) for storing sparse models. In addition, COLMAP provides the option to store the sparse models as text files (human-readable, slow). In both cases, the information is split into three files for the information about cameras, images, and points. Any directory containing those three files constitutes a sparse model. The binary files have the file extension .bin and the text files the file extension .txt. Note that when loading a model from a directory which contains both binary and text files, COLMAP prefers the binary format.
To export the currently selected model in the GUI, choose
File > Export
model. To export all reconstructed models in the current dataset, choose
File > Export all. The selected folder then contains the three files, and
for convenience, the current project configuration for importing the model to
COLMAP. To import the exported models, e.g., for visualization or to resume the
File > Import model and select the folder containing
the cameras, images, and points3D files.
To convert between the binary and text format in the GUI, you can load the model
File > Import model and then export the model in the desired output
File > Export model (binary) or
File > Export model as text
(text). In addition, you can export sparse models to other formats, such as
VisualSfM’s NVM, Bundler files, PLY, VRML, etc., using
File > Export as....
To convert between various formats from the CLI, use the
There are two source files to conveniently read the sparse reconstructions using
scripts/python/read_model.py supporting binary and text) and Matlab
scripts/matlab/read_model.m supporting text).
COLMAP exports the following three text files for every reconstructed model: cameras.txt, images.txt, and points3D.txt. Comments start with a leading “#” character and are ignored. The first comment lines briefly describe the format of the text files, as described in more detailed on this page.
This file contains the intrinsic parameters of all reconstructed cameras in the dataset using one line per camera, e.g.:
# Camera list with one line of data per camera: # CAMERA_ID, MODEL, WIDTH, HEIGHT, PARAMS # Number of cameras: 3 1 SIMPLE_PINHOLE 3072 2304 2559.81 1536 1152 2 PINHOLE 3072 2304 2560.56 2560.56 1536 1152 3 SIMPLE_RADIAL 3072 2304 2559.69 1536 1152 -0.0218531
Here, the dataset contains 3 cameras based using different distortion models with the same sensor dimensions (width: 3072, height: 2304). The length of parameters is variable and depends on the camera model. For the first camera, there are 3 parameters with a single focal length of 2559.81 pixels and a principal point at pixel location (1536, 1152). The intrinsic parameters of a camera can be shared by multiple images, which refer to cameras using the unique identifier CAMERA_ID.
This file contains the pose and keypoints of all reconstructed images in the dataset using two lines per image, e.g.:
# Image list with two lines of data per image: # IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME # POINTS2D as (X, Y, POINT3D_ID) # Number of images: 2, mean observations per image: 2 1 0.851773 0.0165051 0.503764 -0.142941 -0.737434 1.02973 3.74354 1 P1180141.JPG 2362.39 248.498 58396 1784.7 268.254 59027 1784.7 268.254 -1 2 0.851773 0.0165051 0.503764 -0.142941 -0.737434 1.02973 3.74354 1 P1180142.JPG 1190.83 663.957 23056 1258.77 640.354 59070
Here, the first two lines define the information of the first image, and so on.
The reconstructed pose of an image is specified as the projection from world to
the camera coordinate system of an image using a quaternion (QW, QX, QY, QZ)
and a translation vector (TX, TY, TZ). The quaternion is defined using the
Hamilton convention, which is, for example, also used by the Eigen library. The
coordinates of the projection/camera center are given by
-R^t * T, where
R^t is the inverse/transpose of the 3x3 rotation matrix composed from the
T is the translation vector. The local camera coordinate
system of an image is defined in a way that the X axis points to the right, the
Y axis to the bottom, and the Z axis to the front as seen from the image.
Both images in the example above use the same camera model and share intrinsics (CAMERA_ID = 1). The image name is relative to the selected base image folder of the project. The first image has 3 keypoints and the second image has 2 keypoints, while the location of the keypoints is specified in pixel coordinates. Both images observe 2 3D points and note that the last keypoint of the first image does not observe a 3D point in the reconstruction as the 3D point identifier is -1.
This file contains the information of all reconstructed 3D points in the dataset using one line per point, e.g.:
# 3D point list with one line of data per point: # POINT3D_ID, X, Y, Z, R, G, B, ERROR, TRACK as (IMAGE_ID, POINT2D_IDX) # Number of points: 3, mean track length: 3.3334 63390 1.67241 0.292931 0.609726 115 121 122 1.33927 16 6542 15 7345 6 6714 14 7227 63376 2.01848 0.108877 -0.0260841 102 209 250 1.73449 16 6519 15 7322 14 7212 8 3991 63371 1.71102 0.28566 0.53475 245 251 249 0.612829 118 4140 117 4473
Here, there are three reconstructed 3D points, where POINT2D_IDX defines the zero-based index of the keypoint in the images.txt file. The error is given in pixels of reprojection error and is only updated after global bundle adjustment.
COLMAP uses the following workspace folder structure:
+── images │ +── image1.jpg │ +── image2.jpg │ +── ... +── sparse │ +── cameras.txt │ +── images.txt │ +── points3D.txt +── stereo │ +── consistency_graphs │ │ +── image1.jpg.photometric.bin │ │ +── image2.jpg.photometric.bin │ │ +── ... │ +── depth_maps │ │ +── image1.jpg.photometric.bin │ │ +── image2.jpg.photometric.bin │ │ +── ... │ +── normal_maps │ │ +── image1.jpg.photometric.bin │ │ +── image2.jpg.photometric.bin │ │ +── ... │ +── patch-match.cfg │ +── fusion.cfg +── fused.ply +── meshed-poisson.ply +── meshed-delaunay.ply +── run-colmap-geometric.sh +── run-colmap-photometric.sh
Here, the images folder contains the undistorted images, the sparse folder contains the sparse reconstruction with undistorted cameras, the stereo folder contains the stereo reconstruction results, point-cloud.ply and mesh.ply are the results of the fusion and meshing procedure, and run-colmap-geometric.sh and run-colmap-photometric.sh contain example command-line usage to perform the dense reconstruction.
Depth and Normal Maps¶
The depth maps are stored as mixed text and binary files. The text header
defines the dimensions of the image in the format
followed by row-major float32 binary data. For depth maps
for normal maps
channels=3. The depth and normal maps can be conveniently
read with Python using the functions in
with Matlab using the functions in
The consistency graph defines, for all pixels in an image, the source images a
pixel is consistent with. The graph is stored as a mixed text and binary file,
while the text part is equivalent to the depth and normal maps and the binary
part is a continuous list of int32 values in the format
(row, col) defines the
location of the pixel in the image followed by a list of
N image indices.
The indices are specified w.r.t. the ordering in the