Disparity between Eq.(18) in the paper and code implementation

[cailile]

Dear authors, I notice there may be some disparities between Eq.(18) and implementation:

1. In robust_loss.py, Ln.92, the robust regression loss is implemented as:

Compared to what is defined in Eq.(18) of the paper:

there is an extra power term **2 on the cs term.

2. The paper mentioned that c is chosen to be 0.03, but in train_roma_outdoor.py, Ln.220, the value passed to c is 1e-4:

3. The first minus sign in Eq.(18) (indicated by the red circle above) should be removed.

Could you kindly help to clarify? Thanks!

[Parskatt]

1.(and 3) You are correct, eq. 16 is correct but forgot to flip the sign for the KL divergence, thanks.
2. The value we mention in the paper is in pixels, while the codebase is in normalized coordinates. At a resolution of 560 we get $560/2 10^{-4} \approx 0.03 $