Large Depth Error

[codyreading]

Hi, and thanks for the great work!

I have reproduced the experiments in Table 1, and for the most part the metrics are very similar to reported in the paper. The only inconsistency I found was the Depth metric.

Paper:

Reproduced:

I was wondering if this type of behavior is expected between different runs?

[ethanweber]

Hi @codyreading, that's interesting. How do the depth maps look qualitatively? Does that also look incorrect?

[codyreading]

Depth Maps look quite good qualitatively, I've uploaded some of the results here of nerfacto-visibility-cube:

If you could upload your depth maps, I'd love to compare them to visually see any differences.

[ethanweber]

Hi @codyreading, I requested access to your link. When you compute the metrics, do you clip the depth from 0 to 2? We do that in our evaluation here.

[ethanweber]

The depth maps you have look correct as far as I can tell. One more thing to double check is that you are using median depth instead of expected depth when rendering.

[codyreading]

Okay let me check these out, thanks!

[codyreading]

Hey @ethanweber! Apologies for the delay, but just checked that we clipped the depth from 0-2 and also used median depth, I kept all the defaults.

I can rerun this to see if this type of behaviour is reoccuring.

[codyreading]

Closing for now, I don't need to reproduce the depth metrics any longer