Add average_init_density to improve robustness of nerfacto training (#2834)

This commit is contained in:
J.Y authored and GitHub committed 2024-01-28 14:09:27 -08:00
1 parent fb1fee1d8f
commit d8b517d6a5
4 files changed
+14 -2

No files matched your search

+3 -1
View File
@@ -56,12 +56,14 @@ class HashMLPDensityField(Field):
base_res: int = 16,
log2_hashmap_size: int = 18,
features_per_level: int = 2,
average_init_density: float = 1.0,
implementation: Literal["tcnn", "torch"] = "tcnn",
) -> None:
super().__init__()
self.register_buffer("aabb", aabb)
self.spatial_distortion = spatial_distortion
self.use_linear = use_linear
self.average_init_density = average_init_density
self.register_buffer("max_res", torch.tensor(max_res))
self.register_buffer("num_levels", torch.tensor(num_levels))
@@ -111,7 +113,7 @@ class HashMLPDensityField(Field):
# Rectifying the density with an exponential is much more stable than a ReLU or
# softplus, because it enables high post-activation (float32) density outputs
# from smaller internal (float16) parameters.
density = trunc_exp(density_before_activation)
density = self.average_init_density * trunc_exp(density_before_activation)
density = density * selector[..., None]
return density, None
+3 -1
View File
@@ -95,6 +95,7 @@ class NerfactoField(Field):
use_pred_normals: bool = False,
use_average_appearance_embedding: bool = False,
spatial_distortion: Optional[SpatialDistortion] = None,
average_init_density: float = 1.0,
implementation: Literal["tcnn", "torch"] = "tcnn",
) -> None:
super().__init__()
@@ -116,6 +117,7 @@ class NerfactoField(Field):
self.use_pred_normals = use_pred_normals
self.pass_semantic_gradients = pass_semantic_gradients
self.base_res = base_res
self.average_init_density = average_init_density
self.step = 0
self.direction_encoding = SHEncoding(
@@ -218,7 +220,7 @@ class NerfactoField(Field):
# Rectifying the density with an exponential is much more stable than a ReLU or
# softplus, because it enables high post-activation (float32) density outputs
# from smaller internal (float16) parameters.
density = trunc_exp(density_before_activation.to(positions))
density = self.average_init_density * trunc_exp(density_before_activation.to(positions))
density = density * selector[..., None]
return density, base_mlp_out