Professional Documents
Culture Documents
L ab A ssignment
Stereo Reconstruction
May 2, 2012
Timing
Workload : 16 hours 2 preparatory hours 4 Lab hours (programming / solving doubts) 4 programming hours 6 hours for reporting the work done
Data Set
The Data Set provided contains all images necessary for a stereo camera set up calibration, hence no actual images must be acquired. A total of 7 images for the left camera and 7 more for the right one containing a checker board in different positions are provided (see gure 1). In addition, a Matlab script test_triangulacio.m is provided so that you can test how good is your calibration by reconstructing a number of point-and-click points on provided object images.
Stereo Calibration
The B ouguet Matlab toolbox is provided in the dataset as well. Please refer to http://www.vision.caltech.edu/bouguetj/calib_doc/index.html for a thorough description and go-through examples to get acquainted with such toolbox. Please note the importance of obtaining good knowledge of how this toolbox works for the succesful completion of your activity.
The calibration data generated by the Bouguet toolbox can be saved in a .mat le which you can later recover.
1
for right cameras, where 4 corners of a box are being clicked. This is what the script mentioned above can do. Please note that the points clicked in the left and right data sets must correspond. Figure 3 shows the reconstruction of the clicked points in the box. A total of 4 objects are provided in the form of left and right images, namely obj_left_1.jpg, obj_left_2.jpg, etc, and their corresponding _right-ed counterparts. By default, the script opens the object 1 les, so modify the script accordingly if you wish to test with more objects.
Find Correspondences
F or at least two of the object images provided, obtain a representative number (at least 6) of point features using the point-and-click utility provided. O ptionally, in addition to the above point-and-click method, you may automatically select a signicant number of point features using SIFT for both the left and right images, and use Epipolar Geometry in order to obtain point matchings between both the left and right cameras. For either the rst or both methods mentioned above, clearly show the correspondences by displaying the left and right images side by side, with their corresponding point features overlayed, and draw
Figure 3: A 3D view of the reconstructed 4 corners.
lines joining the correspondences. If necessary, apply one of the learned outlier rejection techniques in order to get rid of them.
3D Reconstruction
A pply triangulation in order to obtain a 3D reconstruction of the point correspondences obtained. A short note by Bouguet in le triangulation.ps is provided with the dataset, that shows a possible triangulation technique to be applied. In addition, the Bouguet toolbox provides a function script stereo_triangulation.m that does this job, please refer to the help documentation provided with the toolbox. Try to obtain 3D reconstruction using both the method explained in the short note and the script provided in the toolbox.
Deadline
S unday, 20th M ay 2012 Note: Before the deadline send an email to josep.forest@gmx.com including: - Your name and family name - The matlab le(s) indicating the main le, if necessary - The write-up of the assignment (only PDF) (should not exceed 8 pages in 12 point size font) including: Explain how you have solved every step Include the .m code that solve each step Include the obtained results per step Discuss, if necessary, the results obtained In the end of the document include general comments suggestions regarding the adequacy of the exercise, difculty, workload, etc.