MAE computation is different from Ref-NeRF about nmf HOT 6 CLOSED

Thanks for your update.

I have encountered another problem when reproducing the experiment on helmet. I run the experiment with:
python train.py -m expname=v38_noupsample model=microfacet_tensorf2 dataset=helmet vis_every=5000 datadir={dataset dir}

But it seems to fail to learn the correct geometry and normal vector.

I'm not very clear on whether the novel view synthetic results in Table 1 were trained on HDR or original images from Shiny Blender. Would the use of HDR images influence the results?

from nmf.

Thanks for bringing this to my attention. I will recalculate the values and update the paper.

New values:
blender_dataset norm_err: 20.952093958854675
shiny_dataset norm_err: 6.060845931371053
(updated)

I'll also fix the NVDiffRec and NVDiffRecMC values.

from nmf.

I have updated the arxiv paper.

from nmf.

This is probably caused by the incorrect mixing mode. Can you check the output config.yaml to see if the diffuse mixing mode is set to "fresnel"?

diffuse_mixing_mode: "fresnel"

If it isn't you can set this by passing model.arch.model.diffuse_mixing_mode="fresnel"

from nmf.

I'm using the latest version of code, where the diffuse_mixing_mode is set "fresnel" by default.

By comparing the output config.yaml with your relighting experiment's config.yaml, I have found that only by setting field.smoothing to 1, can the correct normals be obtained.

The command I use:
python train.py -m expname=smoothing model=microfacet_tensorf2 dataset=helmet field.smoothing=1 vis_every=5000 datadir={dataset dir}

from nmf.

Thanks for the help! I have made this the default.

from nmf.