Eval table

[AdamRashid96]

Added eval table to the docs.

[vye16]

lgtm!

[jb-ye]

Could you describe how do you process the images to generate the undistorted images and colmap project? Are you using Inria's convert.py? @AdamRashid96

[AdamRashid96]

Could you describe how do you process the images to generate the undistorted images and colmap project? Are you using Inria's convert.py? @AdamRashid96

I ran Inria's convert.py for the Mipnerf-360 dataset, and then I generated a transforms.json file with the nerfstudio formatting, so I could train nerfstudio's splatfacto with the same poses and colmap parameters.

[jb-ye]

@AdamRashid96 When you convert data to nerfstudio format, did you use the undistorted or distorted colmap cameras?

One more thing is how train/val images are split. Inria use every 8th image as validation image, while nerfstudio's behavior depends on the data parser. If using colmap parser, it is the same, every 8th image. If using nerfstudio parser, it is 10% of input images.

[jb-ye]

I strongly recommend to re-run splatfacto's evaluation using colmap parser after this fix: #136 properly integrated.

[ichsan2895]

Could you describe how do you process the images to generate the undistorted images and colmap project? Are you using Inria's convert.py? @AdamRashid96

I ran Inria's convert.py for the Mipnerf-360 dataset, and then I generated a transforms.json file with the nerfstudio formatting, so I could train nerfstudio's splatfacto with the same poses and colmap parameters.

Why you choose python3 convert.py instead of nerfstudio-dataparser.

FYI I have implement ns-process-data splatfacto in my own repo ichsan2895/nerfstudio. It mimics python convert.py but it compatibles with nerfstudio-dataparser because it creates sparse_pc.ply and transform.json.