Add average_init_density to improve robustness of nerfacto training

nerfstudio/configs/method_configs.py

@@ -97,6 +97,7 @@
         ),
         model=NerfactoModelConfig(
             eval_num_rays_per_chunk=1 << 15,
+            average_init_density=0.01,
             camera_optimizer=CameraOptimizerConfig(mode="SO3xR3"),
         ),
     ),
@@ -140,6 +141,7 @@
             max_res=4096,
             proposal_weights_anneal_max_num_iters=5000,
             log2_hashmap_size=21,
+            average_init_density=0.01,
             camera_optimizer=CameraOptimizerConfig(mode="SO3xR3"),
         ),
     ),
@@ -187,6 +189,7 @@
             max_res=8192,
             proposal_weights_anneal_max_num_iters=5000,
             log2_hashmap_size=21,
+            average_init_density=0.01,
             camera_optimizer=CameraOptimizerConfig(mode="SO3xR3"),
         ),
     ),

nerfstudio/fields/density_fields.py

@@ -56,12 +56,14 @@ def __init__(
         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 @@ def get_density(self, ray_samples: RaySamples) -> Tuple[Tensor, None]:
         # 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
 

nerfstudio/fields/nerfacto_field.py

@@ -95,6 +95,7 @@ def __init__(
         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 @@ def __init__(
         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 @@ def get_density(self, ray_samples: RaySamples) -> Tuple[Tensor, Tensor]:
         # 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
 

nerfstudio/models/nerfacto.py

@@ -124,6 +124,8 @@ class NerfactoModelConfig(ModelConfig):
     """Which implementation to use for the model."""
     appearance_embed_dim: int = 32
     """Dimension of the appearance embedding."""
+    average_init_density: float = 1.0
+    """Average initial density output from MLP. """
     camera_optimizer: CameraOptimizerConfig = field(default_factory=lambda: CameraOptimizerConfig(mode="SO3xR3"))
     """Config of the camera optimizer to use"""
 
@@ -162,6 +164,7 @@ def populate_modules(self):
             use_pred_normals=self.config.predict_normals,
             use_average_appearance_embedding=self.config.use_average_appearance_embedding,
             appearance_embedding_dim=self.config.appearance_embed_dim,
+            average_init_density=self.config.average_init_density,
             implementation=self.config.implementation,
         )
 
@@ -179,6 +182,7 @@ def populate_modules(self):
                 self.scene_box.aabb,
                 spatial_distortion=scene_contraction,
                 **prop_net_args,
+                average_init_density=self.config.average_init_density,
                 implementation=self.config.implementation,
             )
             self.proposal_networks.append(network)
@@ -190,6 +194,7 @@ def populate_modules(self):
                     self.scene_box.aabb,
                     spatial_distortion=scene_contraction,
                     **prop_net_args,
+                    average_init_density=self.config.average_init_density,
                     implementation=self.config.implementation,
                 )
                 self.proposal_networks.append(network)