Class Project: 3D reconstruction from two unrectified images
Date: May 2009
Based on:
1. Pollefeys et al. Visual modeling with a hand-held camera. IJCV, 2004
2. Csurka et al. Characterizing the uncertainty of the fundamental matrix. CVIU, 1997.
3. Pollefeys et al. A simple and efficient rectification method for general motion. ICCV, 1999.
4. Oram. Rectification for any epipolar geometry. BMVC, 2001.
5. Sturm. On focal length calibration from two views. CVPR, 2001.
6. Zhang and Seitz. Parameter estimation for mrf stereo. CVPR, 2005.

The project is motivated by the pipeline of [1]. I started off my establishing keypoint matching using SIFT and the 8-point algorithm,
then I estimate the fundamental matrix as in [2] using Levenberg-Marquardt. I can compute the homography as in [3] and rectify the images
as in [4]. I autocalibrate as in [5] to recover the camera parameters, and then I compute the disparity map using alpha-expansion.
The alpha-expansion parameters are estimated using the EM approach of [6].

Code is online mainly for my reference. It is licensed under GPLv2 for non-commercial research purposes only.