Units of depth maps not matching units of scene

[iahmedmaher]

Hi,

When I try to get the depth of a pixel using the depth map of the rendering modes RGB+ED or RGB+D, I find a big difference with respect to the splats means. The factor could be that of 4. This is not related to the scene scale or anything since when I trace the code I see that the rasterization function takes as input correct units for the each splat's depth, but the resultant map appears to be scaled down in terms of values. Any idea what is the reason for that? And how to know the scale factor for each scene?

[jongwonjlee]

Hi @iahmedmaher, just following up on this discussion, as I'm encountering the same issue you mentioned.

For a sanity check, I compared the ground-truth depth map of the Synthetic NeRF dataset (e.g., Blender, whose raw-files also need to be scaled as discussed here) with the estimate from gsplat. The estimated scale factor needed to match the depth maps appears to be around 3, though this isn't confirmed yet.

I'm just wondering if you have any updates on the scale factor (you mentioned it might be 4, correct?), or any follow-ups on what might be causing this discrepancy. Thanks!

[iahmedmaher]

Thanks for the info!

Sadly I have nothing to add now.

[jongwonjlee]

Just FYI, after I look back, your scale factor of 4 seems correct on my end as well.

[iahmedmaher]

Just FYI, after I look back, your scale factor of 4 seems correct on my end as well.

Thanks for sharing that, do mean empirically or u found some sort of mathematical need for it?