uniform inverse depth sampling VS uniform depth sampling

[thucz]

Hi! I have a problem with the depth candidates for multi-view matching. I noticed that you used a uniform depth sampling for DTU dataset. If I want to use MVSForm++ in indoor or outdoor scenes, Would it be better to use inverse depth sampling (1 / depth) as the maximum value is always very large?

[maybeLx]

We use inverse sampling in all datasets rather than uniform sampling. Please refer to the inverse_depth in the mvsformer++.json

[ewrfcas]

We have also evaluated the performance of inverse and uniform sampling. As a result, inverse always slightly outperforms uniform in our cases.

[thucz]

I see. Thanks!