Lego mesh with grossy background

[ZirongChan]

Hello, I've managed to run the toy_experiment on the provided Lego data.

The camera poses looks fine to me.

the extracted mesh looks good as well.

But I was wondering, what can I do to rule out the noisy vertices on the top and the background? I assume that the "empty space" in the red box below can not be recovered because of the lack of corresponding input images, namely no observation. Is that right?

Looking forwards to your reply. Thx

[mli0603]

Hi @ZirongChan

Yes, you are absolutely right! The noises come from regions that have no observation at all but still lie within the bounding sphere. You could define a tighter bounding sphere manually to improve this. The current automatic bounding region sphere is designed to be relaxed such that we don't miss objects.

[ZirongChan]

Hi @ZirongChan

Yes, you are absolutely right! The noises come from regions that have no observation at all but still lie within the bounding sphere. You could define a tighter bounding sphere manually to improve this. The current automatic bounding region sphere is designed to be relaxed such that we don't miss objects.

@mli0603 thanks for your reply. But why do the noises look different? I mean that the noise above is quite grossly distributed, while the one below the lego platform is very smooth. How do the given camera poses influence this?

[mli0603]

Hi @ZirongChan

There are two types of "noises"

• The top regions that are never observed. These are the regions that simply never correspond to any image observation. Due to the collision of hash encoding, there will be noisy artifacts (looking like small blobs).
• The bottom regions that are observed but ambiguous. These regions are entirely white, and any surface reconstruction can be correct. Due to the observation and our curvature/eikonal losses, they will look like smooth curved surfaces.