Does training in 2k make a significant difference?

[pablodawson]

Hey, thanks for your great work. I noticed you hard-coded a "imageW" and "imageH" value, which it is not actually used afterwards, so the image used is of the original resolution. https://github.com/oppo-us-research/SpacetimeGaussians/blob/72d92222453b590d4eedf9b09504e971d42768fc/thirdparty/gaussian_splatting/scene/dataset_readers.py#L110 I managed to train Neu3D with significantly less (8gb) RAM by resizing the image to that size. Does that make a big difference in your experiments? Thanks.

[lizhan17]

we haven't tried it. because previous methods work on half resolution on Neu3D. But I think 2K resolution will produce a sharper visual. you can see the difference between our demo of n3d and techni color

• n3d is trained with half resoluton.
• technicolor is original resolution. you can change these two dataset with the arguement. "resolution"
• immersive distorted is a specialy case, which we resize the ground truth image directly in the loadercode . To compensate for the resampling error, our method render at the original resolution, that is why resolution for immersive is 1. But the ground truth image for training is half resolution, which is same with other methods.

you can chage resolution by resolution from 2 to 4 in Neu3D, which means only 25%H and 25%W at resolution. this is implemented by 3dgs. please ignore imageH.

[pablodawson]

Got it, thanks! In case you're curious quality seems to be pretty similar.