0

I'd like someone to kindly explain to me how square windows and rectangular windows may affect the computation of the disparity map, and here I'm talking about a fixed size.

I have two images of this object taken from SIR (Random dot pattern on it) enter image description here

I've computed the disparity using 4 x 16 window (MSE = 0.1167), 16 x 4 window (MSE = 0.1333) and 8 x 8 window (MSE = 0.1216). (I could not put the images).

I need to know how can I judge whether the vertical or the horizontal rectangular window is better than its compatible square windows. Is MSE enough for that? Should I also consider the features of the images?

Also, I'd really appreciate it if someone guided me to reconstructing the 3D model.

Thanks in advance.

user3691200
  • 1
  • 3

1 Answers1

2

The window size is a trade off.

The larger you windows are the higher the chances that they will contain useful texture features to provide a good disparity estimate. If they are small they would be more likely to contain smooth areas that do not provide meaningful information.

OTOH, the larger the window is, the more windows you have that contain depth discontinuities. These are the image edges due to multiple objects at different depths and where parallax causes one image to contain pixels occluded in the other and make the window contain multiple disparity values (in different sub regions). These are are the interesting places, but where the block-matching approach actually fails. If your windows are small fewer of them will contain mixed depths giving a more stable results.

Adi Shavit
  • 16,743
  • 5
  • 67
  • 137
  • Great answer. And that eventually will affect the quality of the reconstructed 3D model, right? – user3691200 Feb 12 '17 at 23:34
  • Yes. Block matching is great for smoothly changing surfaces with no depth discontinuities. Since these are quite rear in the real world, other techniques are needed to overcome these problems. – Adi Shavit Feb 13 '17 at 06:38