What is the warmup stage doing exactly?

[yashkant]

Hi, thanks for releasing the code!

I attempted training on the human video, and saw that nerf_coarse is trained for 5 epochs to warmup using an SMPL mesh loaded from mesh_material folder.

However, when I visualized it from tmp/smpl_27554.obj in blender it just seemed to be an ellipsoid [image below], is this expected?

Appreciate your thoughts, thanks!

[gengshan-y]

Yes! It is supposed to be a ellipsoid instead of SMPL shape. As mentioned in "Optimization details" of the preprint -- "We initialize MLP_SDF such that it approximates a unit sphere [65]."

If you open up logdir/$seqname-e120-b256-init/mesh_rest-00.obj to look at the initial mesh at iteration 0, you'll find the following mesh

Note it is a surface extracted from the warmup-ed MLP_SDF by running marching cubes.

[yashkant]

Thanks for the prompt clarification!

I was a bit surprised to find an ellipsoid there and wanted to confirm in case I had introduced a bug somewhere.

[zhan-xu]

Hi! When I tried to train on the human video, it reported "No such file or directory: 'xxx/banmo/tmp/smpl_27554.obj'". Could you let me know how you got this small_27554.obj? Thank you! @yashkant @gengshan-y

[gengshan-y]

Hi, where did it report this error? This smpl_27554.obj file is created here for visualization.

[zhan-xu]

Hi @gengshan-y thanks for the quick reply. I ran the following command

seqname=adult7
python preprocess/img2lines.py --seqname $seqname
bash scripts/template-accu.sh 0 $seqname 10001 "" ""

scripts/template.sh gave the same error.

[gengshan-y]

Could you attach an error log? Is it possible there is no tmp/ dir so that it is not able to save files to it?

[zhan-xu]

Thanks @gengshan-y. That is exactly the problem. The error is gone when I first 'mkdir tmp'

I have another problem. When I tried the pertained human model, I first downloaded human-cap.npy and human-cap.pth, and put them into the tmp folder (as well as opts.log?). I then ran the following commands:

seqname=adult7
bash scripts/render_mgpu.sh 0 $seqname tmp/human-cap.pth "0 1 2 3 4 5 6 7 8 9" 256

Is this the right way? I got however the output as attached, which seems not right. Could you please take a look? adult7_pretrain_output.txt

[gengshan-y]

The problem might be the inconsistent flags stored in the opts.log file. You want to change --nouse_human to --use_human. Let me know if there is still a problem.

[zhan-xu]

Thanks @gengshan-y. That solves the problem. Will let you know if there is further issue. So far things seem work well. Appreciate for your quick help again.