updating theta

[KazukiSakai]

Hi, all.

I'd like to know about optimization algorithm.
In DTAM paper, the following equation are used.

[1] theta^(n+1) = theta^(n) * (1 - beta * n)

But in OpenDTAM another equations are used.

[2] theta^(n+1) = theta^(n) * thetaStep

Why don't you follow original paper?
Is our method better than original one?

And does this difference affects final depth map quality(that is, do both methods make same results after convergence?)

---

Here is my simulation result.
I simulated Eq.[1] with two patterns(beta = 0.001, 0.0001) until n = 1000.
I think finally both behaviors looks same, but when optimization starts, these behaviors are different.
I think you implemented Eq.[2] to achieve faster convergence.

[swordjjy]

Hi, Kazuki

I have no idea how to get in touch with you. So I left this message here. I'm new to DTAM and I have some trouble with running OpenDTAM in my computer. I found that you run the program well, if possible
could give me some tips with how to install the whole stuff? I followed the Instructions and it still not works. I'm using ubuntu 15.XX right now, it is a big problem? Are you using ubuntu 12.04 LTS also in your computer?

[anuranbaka]

Re the original comment, I believe the original idea was to allow any arbitrary shape of the thetastep function, which in the paper is 1-beta*n. I found the exact form didn't seem to matter too much, but using a constant gave a bit more time for the waves to settle in the theta=0.6 region. But I agree it is a bug that the demo in the repo doesn't use the paper's version.

Re Kazuki, I think there are several similar comments in other issues. Mainly check that the GPU is Kepler or higher and that CUDA is compiled into opencv correctly.