A quiz on the stereo experiment

Adrian F. Clark

1. What is the main purpose of this experiment?

   • to see how to match corners

   • to find out how accurate the camera calibration functionality in OpenCV is

   • to explore the use of computer vision on 3D problems
   • to allow us to practise how error (uncertainty) propagation

   The real reason for this experiment is to ascertain whether the ability to calibrate the camera leads to accurate 3D reconstruction. The technique is very widely used in the games and movie industries for capturing human motion, and also finds important application in the medical world for capturing body shape for making prosethics. It is, of course, used in vision-guided robotics too.

2. How is the error (uncertainty) in the focal length estimated by calibrate.py?

   • from the regularity of the spacing of detected corners

   • the difference between horizontal and vertical focal lengths

   • from the accuracy of the located corners
   • from the locations of the corners

   The code in the program uses the difference between the horizontal and vertical focal lengths as an estimate of the uncertainty.

3. How were the synthetic images of Candide created?

   • using Unity
   • using OpenGL
   • using Blender

   • using POV-ray

   The Candide model is rendered using POV-ray, a freely-available, high-quality rendering package based around ray-tracing technology. It is far from being real-time but does (or rather can, for we wouldn't describe the appearance of Candide as being photo-realistic) produce high-quality output. Its most important characteristic for computer vision is that it is driven by a kind of programming language, so one can write scripts to generate POV-ray input files, which is what was done here.

4. How close is the reconstruction by camera calibration to the correct result?

   • nothing like correct
   • exactly correct
   • within two times the error bounds

   • well within the error bounds

   My own calculation from measuring the nose position in the two images was 485.33 and the calibration yielded $(484.66 \pm 1.31)$. The two agree to well within the error (uncertainty) from the calibration, and that will be much less than the uncertainty in the estimation of the nose position.

5. Why does the chessboard pattern have different numbers of squares in the x- and y-directions?

   • to ensure there are enough corners
   • there is no particular reason
   • so that it prints on a sheet of A4 paper

   • to remove any ambiguities over its orientation

   The number of squares differ so that the software can determine that the target has been rotated.

6. What is the maximum number of iterations allowed in calibrate.py?

   • 50
   • 100

   • 30

   • 10

   The definition of term_crit permits up to 30 iterations, admittedly a fairly arbitrary choice.

7. What is estimated by the camera calibration process?

   • the focal length of the camera

   • the baseline of the stereo imaging system
   • the disparity
   • the spacing between pixels

   Camera calibration, at least as used here, estimates the focal length.

8. When calculating the position of Candide's nose in 3D, what contributes most to its error (uncertainty)?

   • the pixel spacing

   • position of her nose

   • the focal length
   • the baseline

   The error (uncertainty) in estimating the position of her nose is by far the least accurate quantity measured, so this is largely what influences the accuracy of the overall result.

9. What is the pattern on a calibration target?

   • a set of plus-shaped lines
   • a QR code

   • a chessboard

   • a set of concentric circles

   Camera calibration usually makes use of a chessboard pattern, though the number of black and white squares does not necessarily match the number on a chessboard. It is used because it is fairly easy to detect and the edges between squares allow sub-pixel accuracy to be found.

10. How does calibrate.py process the images?

    • in RGB space
    • in BGR space
    • in HSV space

    • as grey-scale ones

    The program converts input images to grey-scale before passing them to the routine to find corners.