nerfacc how to visualize the train step

hello , I want to visualize the predicted image when training to show the efficiency of ngp. Is there any relevant settings in the code ?

Oct 27 '22 13:10 liwenssss

You can uncomment here:

https://github.com/KAIR-BAIR/nerfacc/blob/107d7466ef3fcfadd64e8b57ad3c1aafd4874fe0/examples/train_ngp_nerf.py#L298

Oct 27 '22 18:10 liruilong940607

thanks for your reply. Is it possible to use nerfacc to speed up human-nerf (such as animatable_nerf) training? I tried to embed nerfacc in it, but found that some of the Settings were hard to align..

Oct 29 '22 09:10 liwenssss

You can refer to the d-nerf example. human-nerfs are very much like that. You just need to have a more customized warping field conditioned on human pose instead of timestamps.

Checkout the explanation of the d-nerf usage here: https://www.nerfacc.com/en/latest/examples/dnerf.html

Oct 30 '22 00:10 liruilong940607

I tried but got empty ray_indices when eval. It's weird. Do you know what might have caused it. Or do you have plan to give an example on zju dataset?

Nov 01 '22 02:11 liwenssss

Can't tell where go wrong. ray_indices is empty means either an all-zero occupancy grid, or improper bounding-box. You can debug by tracing a single sample where you are certain it should not be zero density, and see at which step it was throw away.

Having an example for human-nerfs would be nice as I also work in this field (checkout my latest paper here: https://github.com/facebookresearch/tava), but there are many other TODOs have higher priority than this.

If you make it work please lmk and I'm happy to accept PR / link to your repo.

Nov 02 '22 10:11 liruilong940607