A project performed as a part of the Computer Vision (TSBB15) course at Linköping University. The purpose of the project is to implement a computer program capable of, given a set of images and known intrinsic camera parameters, estimate the location of image points in 3D as well as the camera pose. 3D reconstruction is, in computer vision, defined as the process of capturing shapes, locations and/or appearances of real objects. The project goal was to create a 3D-reconstruction application capable of performing the aforementioned tasks as well as estimating the camera trajectory. The application was implemented in C++ mainly using the OpenCV library. https://github.com/Epipolarna/3DReconstruction

2.  Main program and pipeline – Mattias
 2D correspondence extraction – Alexander
 Non-linear optimization – Martin
 Visualization – Gustav

4.  Camera data structure
 2D <-> 3D
 Camera pair
& 2D <-> 3D

5.  Difficult data set
 Outliers
 Long time before results are seen

6.  Chains?

7.  Find feature points, using Harris response
 Minimum relative quality
 Minimum distance
 Calculate descriptor, using SIFT
 Calculate correspondences, using Brute Force

10.  Initial relation between cameras is calculated
from the essential matrix, according to
algorithm described in an epic compendium
by Klas Nordberg.
 Results used as initial guess for PnP

11.  Implemented using the levmar API
 Called three times in one pipeline iteration
 Last step of Gold Standard algorithm
 PnP pose estimation
 Bundle adjustment
 Minimizes squared re-projection error over
specified parameters.

12.  Vector representation
 Support exists in OpenCV: (cv::Rodrigues)
 Allows for unconstrained optimization
 Ambiguity in the representation

13.  Initial guess derived from F (R & t)
 Use known correspondences from previous
view
 Threshold new points on re-projection error
to remove outliers

14.  Same rotation parameterization as the one used
for solving the PnP
 Computation time scales badly with the number
of points and views
 Dinosaur set takes more than ten hours to complete
(30 points per view)
 Slow mainly because
 No sparse pattern is used for Jacobian estimation
 levmar was not built using LAPACK

15.  Plot the estimated cameras
 Plot the 3D points.
 Points colored from original
image data.

16. Very slow for many cameras.
Some drift for the last few cameras.
Good results except for the last 3 cameras.