Question about entropy loss

[dhdkc]

Question about entropy loss

[aakash-chaddha]

The entropy loss they refer to is an unsupervised technique to encourage the gaussian opacity values αg to become binary (fully opaque or fully transparent) during the optimization. It does not require any ground truth density data.

Specifically, they add an entropy term to the loss function that looks like:

H(αg) = -αg log(αg) - (1-αg) log(1-αg)

They compute this for each gaussian's scalar opacity value αg.

This entropy measure will be minimized when αg trends towards 0 or 1. In between values result in higher entropy/uncertainty.

So by minimizing this term for all the gaussian opacities, it encourages the opacities to take on extreme binary 0 or 1 values, rather than partial transparent values.

This aligns with their modeling assumptions for mesh extraction that gaussians should fully represent surface content or be fully discarded from surface consideration based on their opacity.

tldr: the entropy loss provides an unsupervised regularization that pulls the gaussian opacities towards 0 or 1 without needing any ground truth density supervision. It helps refine and simplify the geometry representation.

[Anttwo]

It is exactly as Antarashtriya said, thank you!
In practice, we compute this loss only for the Gaussians located in the field of view of the current camera (during a training iteration).

Best!

[dhdkc]

Thanks for your detailed explanation, now I've understand.