OrthoSfM

Official implementation of A Structure from Motion Pipeline for Orthographic Multi-View Images. This software can accurately estimate poses and points from a set of images under orthographic projection.

Installation

A prebuilt version for windows can be found here: Link

For building, the software requires the following third party libaries to be present:

Ceres (required)
Boost (required)
Eigen3 (required)
OpenCV (required)
matplotlibcpp (optional -> only necessary for testbench)

On Windows we highly recommend using vcpkg (https://vcpkg.io) to install the dependencies.

Build options:

-DENABLE_TESTBENCH=True/False --> For building the testbench application

-DENABLE_CUDA=True/False --> Enables CUDA based SIFT feature detection

Datasets

Full synthetic datasets: Download
Real datasets: Download
Subsets of 3 images per dataset: Download

Example Usage

One example how to use the compiled app is:

orthosfm-app.exe {project folder} {Image Folder} --downscale-factor=2 --overwrite --solver=0

A more detailed overview of the available parameters can be obtained by calling

orthosfm-app.exe --help

OrthoSfM: A simple Structure from Motion (SfM) implementation primarily for orthographic images

Usage: OrthoSfM.exe [project folder] [image folder] --option1 value1 --option2 value2
or OrthoSfM.exe --project-folder [project folder] --image-folder [image folder] --option1 value1 --option2 value2

Allowed Additional options:
--help      produce help message
--project-folder arg      The folder where all temporary files and results should be saved. This folder is required.
--image-folder arg      A folder containing the images that should be used as input. This folder is required.
--calculated-tracks arg      A track file containing previously calculated tracks can be specified to avoid calculating them on the fly
--export-pairwise-tracks      Exports the tracks in a pairwise fashion usefull as input to other sfm implementations
--mask-folder arg      A folder containing a mask per image in the format {imageName}_mask.png
--downscale-factor arg      The number of times the input images are downscaled before using them for calculation
--overwrite      Resets any old project in the same project folder and overwrites it with the new project.
--solver arg      The solver used on the dataset [0 = Quaternion, 1 = EulerHorizontal, 2 = EulerHorizontalVertical, 3 = EulerAll]

Testbench

We provide an application to systematically test the implementation on list of datasets.

orthosfm-testbench.exe [project folder] [dataset folder] [executable path]

Project Structure

The folder structure of the project follows the following sheme:

src/ -> contains all source code
- algortihms -> Implementations of different camera models and reprojection errors
- app -> The main command line executable
- bundle_adjustment -> All code that is necessary to run a bundle adjustment
- cuda_sift -> Code from (https://github.com/Celebrandil/CudaSift) for detecting SIFT features on the GPU
- data_structures -> All data structures that are used as part of a reconstruction
- matching -> Interface and implemntation of feature matching algorithms
- mve -> Partial copy of the source code from (https://github.com/simonfuhrmann/mve) used for robust feature extraction and matching
- sfm -> Contains the main pipeline that combines all other parts
- testbench -> A test application used to systematically evaluate the implemented pipeline
- triangulation -> Code to triangulate 3d point clouds from reconstruction data
- util -> Some utility code that is reused by multiple libraries
resources -> Contains three different models used inside the testbench

daydreamer2023 / OrthoSfM