Questions about performance under large-scale scenes

[DekuLiuTesla]

Thanks for releasing your great work!

I reproduced the paper and get nice results on Tanks and Temples Dataset. However, when I apply RaDeGS on "Building" scene of Mill19 dataset (covering region of hundreds of meters, and lambda_distortion has been set to 1 according to scene scale), it produces poor results compared with vanilla 3DGS. Several test views looks good, but many others are occluded by planes or have lower fidelity.

The performance gains a lot when I disable normal and distortion regularization, but performance still lags from vanilla 3DGS. Is there any insight to generalize RaDeGS better to these large-scale scenes?

[MELANCHOLY828]

感谢您发布您的出色作品！

我复现了这篇论文，并在 Tanks and Temples 数据集上获得了不错的结果。但是，当我将 RaDeGS 应用于 Mill19 数据集的“建筑”场景（覆盖数百米的区域，并且 lambda_distortion 已根据场景比例设置为 1）时，它与 vanilla 3DGS 相比产生了较差的结果。几个测试视图看起来不错，但许多其他视图被平面遮挡或保真度较低。

当我禁用法线和失真正则化时，性能会大幅提升，但性能仍然落后于原始 3DGS。有没有什么见解可以让 RaDeGS 更好地适用于这些大规模场景？

Can you provide the code for rendering and evaluation?