Update NGP related models with nerfacc 0.5.2 #1809
Conversation
|
CUDA_VISIBLE_DEVICES=3 ns-train instant-ngp-bounded --data ./data/nerf_synthetic/lego/transforms_train.json --vis wandb --max-num-iterations 10000
|
|
CUDA_VISIBLE_DEVICES=3 ns-train instant-ngp --data data/nerfstudio/poster --vis wandb --max-num-iterations 10000
|
|
CUDA_VISIBLE_DEVICES=3 ns-train nerfplayer-ngp --vis wandb --max-num-iterations 10000 --data ./data/mochi-high-five/
|
|
In summary:
The major difference is that now we re-enable the non-gradient skipping for the occluded regions in the sampling procedure. |
| if self.config.disable_scene_contraction: | ||
| scene_contraction = None | ||
| else: | ||
| scene_contraction = SceneContraction(order=float("inf")) |
There was a problem hiding this comment.
@liruilong940607 is scene contraction meant to be used with the new multi-level grid? Should TCNNNerfactoField just not take the expanded aabb as in https://github.com/KAIR-BAIR/nerfacc/blob/master/examples/train_ngp_nerf_occ.py#L130
There was a problem hiding this comment.
The field and multi-level occ grid are completely separated with each other. The new occ grid does not support contraction anymore but the field can still do contraction. I didn't play around with the field to compare it with/without contraction so I just kept it as is.
Plz share if you find anything!
This PR separates nerfacc update from the PR #1705. It bumps the
nerfaccdependence to its latest version (0.5.2), and updates NGP related models (instant-ngp,instant-ngp-bounded,nerfplayer-ngp) with the latest nerfacc.Note on the master branch, for some reason the non-gradient skipping is disabled with nerfacc. This PR re-enables that with the latest nerfacc.
After this PR, we can update the docs for the ngp-based methods, such as #1774.