Structure-from-Motion for MAV image sequences
C++ Objective-C C
Switch branches/tags
Nothing to show
Clone or download

====== MAVMAP


MAVMAP is a structure-from-motion system.

The system is intended to take a sequence of images (taken from an arbitrary number of cameras and different camera models) as input and produce a 3-D reconstruction of the camera poses and the sparse scene geometry as output.

More information about the internals of the reconstruction process can be found below.

If you find this project useful, please consider citing

Schönberger, J. L., Fraundorfer, F., and Frahm, J.-M.:
Structure-from-motion for MAV image sequence analysis with photogrammetric applications,
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-3, 305-312,
doi:10.5194/isprsarchives-XL-3-305-2014, 2014.


  • Boost
  • Eigen3
  • Ceres-Solver
  • OpenCV

Confirmed to work with the following versions (2013-12-09):

  • Boost==1.55.0
  • Eigen==3.2.0
  • Ceres-Solver=={1.7.0, 1.8.0}
  • OpenCV==
  • Clang==LLVM version 5.0 (clang-500.2.79), GCC==4.6.3

Additionally, it is fully functional on Ubuntu 12.04 LTS with system packages and OpenCV>=2.4.X.

Currently, the system only works on 64bit platforms, but it also runs on 32bit platforms, however with reduced performance, by adding the following compile flag EIGEN_DONT_ALIGN_STATICALLY.


  • Install Ceres-Solver:
    tar zxf ceres-solver-X.X.X.tar.gz
    cd ceres-solver-X.X.X
    mkdir build
    cd build
    cmake ..
    sudo make install
  • Compile MAVMAP:
    cd mavmap
    mkdir build
    cd build
    cmake ..


Free for academic use. For information on commercial licensing, please contact the authors.

Johannes L. Schönberger <johannes.schoenberger (at)>
Friedrich Fraundorfer <friedrich.fraundorfer (at)>


A fully-functional mapping routine is included, which is aimed at reconstructing sequential sets of images, such as image sequences from UAVs.

Image data format

This routine is implemented in src/mapper and can be configured via its command line interface. The input data consists of a set of images whose properties are defined in the imagedata.txt file. All the files must reside in the same directory, e.g.


The imagedata.txt defines the ordering of the acquisition of the images. Consecutive images should optimally have overlaps. Every image is defined in a separate line with some meta-data and the camera model, e.g.

image1, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 1, PINHOLE, 100.000, 100.000, 368.000, 256.000
image2, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5
image3, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5
image4, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 2, OPENCV, 100.000, 100.000, 368.000, 256.000, 0.1, 0.2, 0.01, 0.02
image5, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 1
image6, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5
image7, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5
image8, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 2
image9, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 2
image10, 1.1, -0.3, -1.0, 47.4, 9.2, 485.4, 2.8, 20.3, -0.4, -4.5, 2, OPENCV, 100.000, 100.000, 368.000, 256.000, 0.1, 0.2, 0.01, 0.02
  • {BASENAME}: Filename of image without file extension.
  • {ROLL, PITCH, YAW}: Camera orientation in radians (using the convention R = R_x(roll) * R_y(pitch) * R_z(yaw)) in the world coordinate system.
  • {LAT, LON, ALT}: Global camera position from e.g. a GPS-sensor.
  • {LOCAL_HEIGHT}: Local altitude above ground from e.g. a proximity sensor.
  • {TX, TY, TZ}: Camera position in the world coordinate system.
  • {CAM_IDX, CAM_MODEL, CAM_PARAMS[N]}: Camera index, camera model and camera parameters. Each unique camera must get an unique camera index. The camera model is assumed to be equal for all the following images until a new camera model with a new index is defined. In the example above this means that image{1,2,3,5,6,7} were taken from the first and image{4,8,9,10} from the second camera with different camera models, respectively. Available camera models and their respective parameter ordering can be found in src/base3d/camera_models.h.

Apart from the image BASENAME only {ROLL, PITCH, YAW} are currently used as input for the optional constraining of the camera poses from e.g. IMU-sensors. The values {ROLL, PITCH, YAW, TX, TY, TZ} will be set in the output data file imagedataout.txt.

Mappers for arbitrary image configurations can be easily built on top of the core components of the library. See internals section below.

Control points and geo-registration

Definition of control points provides the following two features:

  • Geo-registration w.r.t. coordinate system of fixed control points (ground control points)
  • Estimation of specific points by defining variable control points

You must enable --use-control-points and set --control-point-data-path to point at a file in the following format:

  • {FIXED_CONTROL_POINT_NAME,VARIABLE_CONTROL_POINT_NAME}: Unique control point identifier. One # indicates a variable and two # a fixed control point.
  • {X, Y, Z}: Position of control point, use dummy values for variable control points.
  • {IMAGE_IDX}: 0-based index of image in imagedata.txt of observation.
  • {IX, IY}: Pixel position of observation in image.


MAVMAP is built on top of the following core components:

  • FeatureManager: Manage 2-D image feature points, 3-D scene points, camera poses, camera models, 2-D feature correspondences and 3-D point tracks.

  • FeatureCache: Extract feature (SURF) locations and descriptors and cache them on disk for fast later retrieval. Automatic detection of changed feature detection and extraction parameters.

  • SequentialMapper Provide tools for the sequential processing of images, bundle adjustment of the scene and cameras, automatic loop-detection and merging of separate, overlapping scenes.

The SequentialMapper makes use of the FeatureCache for feature retrieval and of FeatureManager for storing the complete reconstructed information (camera poses, scene, feature locations etc.) for efficient memory usage and fast data access. The FeatureManager is a self-contained class without any external dependencies.

The src/mapper routine is intended for the reconstruction of sequential image sets, where each image was sequentially taken one after another. The methods of the SequentialMapper allow for the creation of mapping routines of arbitrary image collections.


The data structures in the class were specifically chosen for fast access, convenient usage and low memory footprint. The data is generally guaranteed to have constant access and insertion time.

The entire structure of the scene and the camera poses is stored in the FeatureManager and does not require any additional information for bundle adjustment and data output.

The core functionality of the manager is to keep track of feature correspondences in that it automatically creates new 3-D points, merges existing tracks, eliminates duplicate observations of the same 3-D point in the same image.

Data access to the std::unordered_map containers should be careful. While std::unordered_map[] is the fastest access, it automatically creates a new key-value pair in the map if the key does not exist. It can destroy the automatic track list functionality. It is thus recommended to use std::unordered_map::at for data access from outside.


This class is intended for the following exemplary process chain:

  1. Find good initial image pair and reconstruct structure by estimating the essential matrix. (See process_initial)
  2. Sequentially reconstruct new camera poses and extend the existing scene geometry by processing new images against already reconstructed images or by processing already reconstructed images (but a new combination) against each other to extend and improve the scene geometry. This routine is based on 2D-3D pose estimation by using already reconstructed 3-D world points and their corresponding image features in the new image. (See process)
  3. Adjust local or global bundle. (See adjust_bundle)
  4. Detect loops in already reconstructed scene. (See detect_loop)
  5. Try to merge different sequential mappers with overlapping scene parts. (See merge).

Exemplary usage:

SequentialMapper mapper1(...);
mapper1.process_initial(image1, image2);
mapper1.process(image3, image2);
mapper1.process(image3, image1);
mapper1.process(image4, image3);
mapper1.process(image5, image4);
mapper1.process(image100, image4);

SequentialMapper mapper2(...);
mapper2.process_initial(image1, image2);
mapper2.process(image3, image2);


This class always keeps the previously extracted features in memory and it is thus recommended to sequentially process images for optimum performance, so that the next new image is processed against one of the last images, i.e.

SequentialMapper mapper(...);
mapper.process_initial(image1, image2);
mapper.process(image3, image2);
mapper.process(image4, image3);
mapper.process(image5, image4);
mapper.process(image6, image4);
mapper.process(image6, image5);

More details about the usage of the class can be found in the header source file.

Bundle Adjustment

The bundle adjustment is based on the Ceres-Solver library.

It is possible to combine arbitrary camera models in the same adjustment problem. The solver uses the robust Cauchy loss function and the cost function is based on the reprojection error in pixels. The following convention for the projection of 3-D world (X) to 2-D image (x) points is employed:

T = [R | t]
    [0 | 1]
X' = T * X
x = C(X')

where T is a 4x4 matrix and transforms 3-D points from the world to the camera coordinate system. R is a 3x3 transformation matrix and t is a 3x1 translation vector. R (in angle-axis representation) and t are stored in the FeatureManager. The camera pose in the world coordinate system can be calculated by extracting R' and t' from T^-1.

C(X) represents the camera model and projects the 3-D point in the camera coordinate system to the image plane. For the standard pinhole camera model it is defined as:

C(X) = f * X + c

where f is the focal length and and c the principal point.

The bundle adjuster optimizes the poses and 3-D points of the feature manager in-place, i.e. it uses only very few additional memory for building the bundle adjustment problem.

The solver uses the sparse Schur algorithm for solving the normal equations and the parameter and residual ordering uses Ceres' automatic ordering routine.

Each 3-D point in the bundle adjustment is guaranteed to have at least two observations in separate images to avoid singularities in the Jacobian. At least two poses should be set as fixed to avoid datum defects which result in rank deficient Jacobians.