Structure from Motion using OpenCV

An easy understandle Structure from Motion (SfM) application wirrten in C++ using OpenCV, Qt and OpenGL.

The goal of this application was mainly educational (Bachelor Thesis by nordewal/flec). It gives one a good insight and starting point into the Structure from Motion domain. Using the GUI one can easily follow the pipeline used for SfM:

Detect and describe features of each image
Create image pair and match features
Compute rotation and translation (extrinsic parameters) for the first image pair
Create a point cloud for the first image pair.
Add more points to the point cloud using PnP and triangulation.
Minimize the errors in the point cloud using bundle adjustment.

This application can only handle images that are already in the correct sequence (eg. video), as it only compares image A to B and image B to C. But not image A to C. The application can load images as well as videos.

The main logic is in sfmapp.h and sfmapp.cpp. One can easily interchange feature detectors and so on in the sfmapp.h file.

Installation

CUDA

OPTIONAL: One can offload feature detection to the graphic card. Thus enable the USE_CUDA CMake option and compile OpenCV with CUDA.

Linux

Build OpenCV with the contrib modules and install it. Make after cloning the git repos (opencv & opencv_contrib) that you checkout the same version of both repos:

git clone https://github.com/Itseez/opencv_contrib.git
cd opencv_contrib
git checkout 3.0.0
cd ..
git clone https://github.com/Itseez/opencv.git
cd opencv
git checkout 3.0.0
mkdir build
cd build
cmake -DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules ..

Install the following dependencies:
- libsuitesparse-dev
- Qt (working for us: 5.5.0)
- libqglviewer-dev (working for us: 2.6.3+dfsg1-1)
Compile the code:

mkdir build
cd build
cmake ..
make

Usage

Datasets

For sake of comfort, two datasets are already provided in the datasets folder:

house: Created for this project
fountain: Created by the EPFL: http://cvlabwww.epfl.ch/data/multiview/denseMVS.html

Camera intrinsic parameters

To avoid specifying the camera intrinsic parameters on every run in the GUI, one can define those in a file called camera_intrinsic.yaml, which must be in the same folder as the images. Replace fx, fy, cx and cy accordingly.

%YAML:1.0
mat1: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [fx, 0,  cx, 
          0,  fy, cy,
          0,  0,  1 ]

Camera distorsion parameters

For automatically rectifying the images before they are being processed, one can provide the distorsion parameters of the camera. Thus the file camera_distortion.yaml must be in the same folder as the images. Replace the number of columns (size of parameters; either 4, 5 or 8), and the data array accordingly:

%YAML:1.0
mat1: !!opencv-matrix
   rows: 1
   cols: 8 # number parameters
   dt: d
   data: [k1, k2, p1, p2, k3, k4, k5, k6] # Values of the parameters

nordewal/sfm