I am researching KLT tracking algorithm and I came across this paper. It says:
However, smaller windows are in general preferable for tracking because they are less likely to straddle a depth discontinuity. For this reason, a pure translation model is preferable during tracking,...
I am trying to understand why a separate translation model is needed when the affine model is being calculated anyway. Here are the things that I have observed:
- A KLT implementation can work without using Affine Transformation for calculation of residue, by relying only on the much simpler translation model.
- The window size is different for both models.
- The affine model for the residue calculation need not be very accurate which means lower computation expense.
I think the observations 2 and 3, along with the issue of "straddling a depth discontinuity" provides a good set of reasons to have two separate models.
My questions are:
- Are the above three observations correct?
- What exactly does "straddle a depth discontinuity" mean? I am not a native English speaker and I am not able to connect the meaning of the word 'straddle' with this context.