Training configurations of vanilla 3D-GS

[XuHu0529]

Hi, thanks for your excellent work!

In Table 1 of S32 datasets, the PSNR values of vanilla 3D-GS are about 26 on novel views. I run vanilla 3D-GS with LiDAR points and 3 cameras, but only get the PSNR of about 22.

Could you please tell me the concrete training configurations?

Thank you very much!

[nnanhuang]

We did make some adjustments, such as aligning the parameters and training strategies used in our codebase, and adding LiDAR point cloud initialization and depth supervision. I suspect these changes are the reasons for the higher metrics. We are currently organizing the baseline code and will update it soon.

[ElenaViewSynthesis]

Hey @nnanhuang ,
Could you provide the modified technical configurations on your codebase for obtaining these metrics for a higher-quality reconstruction?

Also can you be more specific about the types of parameters you aligned and the approach you followed for depth supervision?