reproduce results on Tanks&Temples Auditorium

[jinmaodaomaye2021]

Hi author,

Thanks for your contribution of using monocular cues to supervise 3D GS. I am trying to reconstruct the Tanks&Temples Auditorium scene with your code and compare with the results from monosdf. Unfortunately, I found the mesh reconstruction from your code is inferior to monosdf. I am not sure if I followed the same setting as your paper.

The first video is from your code and the second video is from monosdf.

I also attached the rendered depth map and the normal map from DN-Splatter. The depth map looks ok and the normal map is not smooth. Any idea why ?

Thanks.

Experiment setting:

• mono depth supervision with ZoeDepth, aligned with SfM sparse depth

• surface normal supervision with Omni data

• command for training

ns-train dn-splatter --max-num-iterations 30000 --pipeline.model.use-depth-loss True --pipeline.model.mono-depth-lambda 0.2 --pipeline.model.use-depth-smooth-loss True --pipeline.model.use-normal-loss True --pipeline.model.normal-supervision mono coolermap --data Auditorium --depth-mode mono --load-normals True --normals-from pretrained --normal-format opengl --load-depths True

• command for mesh extraction

gs-mesh dn --load-config Auditorium/config.yml --output-dir Auditorium

https://github.com/maturk/dn-splatter/assets/89769436/3486b281-0d9c-442e-bbdc-4050811249a2

https://github.com/maturk/dn-splatter/assets/89769436/e03bb543-ca23-461e-8835-8a885f18fc47