Add flag to disable apperance embedding (#2843)

nerfstudio/fields/nerfacto_field.py

@@ -110,7 +110,8 @@ class NerfactoField(Field):
         self.spatial_distortion = spatial_distortion
         self.num_images = num_images
         self.appearance_embedding_dim = appearance_embedding_dim
-        self.embedding_appearance = Embedding(self.num_images, self.appearance_embedding_dim)
+        if self.appearance_embedding_dim > 0:
+            self.embedding_appearance = Embedding(self.num_images, self.appearance_embedding_dim)
         self.use_average_appearance_embedding = use_average_appearance_embedding
         self.use_transient_embedding = use_transient_embedding
         self.use_semantics = use_semantics
@@ -239,17 +240,19 @@ class NerfactoField(Field):
         outputs_shape = ray_samples.frustums.directions.shape[:-1]
 
         # appearance
-        if self.training:
-            embedded_appearance = self.embedding_appearance(camera_indices)
-        else:
-            if self.use_average_appearance_embedding:
-                embedded_appearance = torch.ones(
-                    (*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
-                ) * self.embedding_appearance.mean(dim=0)
+        embedded_appearance = None
+        if self.appearance_embedding_dim > 0:
+            if self.training:
+                embedded_appearance = self.embedding_appearance(camera_indices)
             else:
-                embedded_appearance = torch.zeros(
-                    (*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
-                )
+                if self.use_average_appearance_embedding:
+                    embedded_appearance = torch.ones(
+                        (*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
+                    ) * self.embedding_appearance.mean(dim=0)
+                else:
+                    embedded_appearance = torch.zeros(
+                        (*directions.shape[:-1], self.appearance_embedding_dim), device=directions.device
+                    )
 
         # transients
         if self.use_transient_embedding and self.training:
@@ -289,8 +292,10 @@ class NerfactoField(Field):
             [
                 d,
                 density_embedding.view(-1, self.geo_feat_dim),
-                embedded_appearance.view(-1, self.appearance_embedding_dim),
-            ],
+            ]
+            + (
+                [embedded_appearance.view(-1, self.appearance_embedding_dim)] if embedded_appearance is not None else []
+            ),
             dim=-1,
         )
         rgb = self.mlp_head(h).view(*outputs_shape, -1).to(directions)

nerfstudio/fields/nerfw_field.py

@@ -75,7 +75,8 @@ class VanillaNerfWField(Field):
         self.appearance_embedding_dim = appearance_embedding_dim
         self.transient_embedding_dim = transient_embedding_dim
 
-        self.embedding_appearance = Embedding(self.num_images, self.appearance_embedding_dim)
+        if self.appearance_embedding_dim > 0:
+            self.embedding_appearance = Embedding(self.num_images, self.appearance_embedding_dim)
         self.embedding_transient = Embedding(self.num_images, self.transient_embedding_dim)
 
         self.mlp_base = MLP(
@@ -96,7 +97,7 @@ class VanillaNerfWField(Field):
         self.mlp_head = MLP(
             in_dim=self.mlp_base.get_out_dim()
             + self.direction_encoding.get_out_dim()
-            + self.embedding_appearance.get_out_dim(),
+            + (self.embedding_appearance.get_out_dim() if self.appearance_embedding_dim > 0 else 0),
             num_layers=head_mlp_num_layers,
             layer_width=head_mlp_layer_width,
             out_activation=nn.ReLU(),
@@ -134,9 +135,11 @@ class VanillaNerfWField(Field):
         if ray_samples.camera_indices is None:
             raise AttributeError("Camera indices are not provided.")
         camera_indices = ray_samples.camera_indices.squeeze().to(ray_samples.frustums.origins.device)
-        embedded_appearance = self.embedding_appearance(camera_indices)
-        mlp_in = torch.cat([density_embedding, encoded_dir, embedded_appearance], dim=-1)  # type: ignore
-        mlp_head_out = self.mlp_head(mlp_in)
+        mlp_in = [density_embedding, encoded_dir]
+        if self.appearance_embedding_dim > 0:
+            embedded_appearance = self.embedding_appearance(camera_indices)
+            mlp_in.append(embedded_appearance)
+        mlp_head_out = self.mlp_head(torch.cat(mlp_in, dim=-1))
         outputs[self.field_head_rgb.field_head_name] = self.field_head_rgb(mlp_head_out)  # static rgb
         embedded_transient = self.embedding_transient(camera_indices)
         transient_mlp_in = torch.cat([density_embedding, embedded_transient], dim=-1)  # type: ignore

nerfstudio/models/nerfacto.py

@@ -106,6 +106,8 @@ class NerfactoModelConfig(ModelConfig):
     """Predicted normal loss multiplier."""
     use_proposal_weight_anneal: bool = True
     """Whether to use proposal weight annealing."""
+    use_appearance_embedding: bool = True
+    """Whether to use an appearance embedding."""
     use_average_appearance_embedding: bool = True
     """Whether to use average appearance embedding or zeros for inference."""
     proposal_weights_anneal_slope: float = 10.0
@@ -148,6 +150,8 @@ class NerfactoModel(Model):
         else:
             scene_contraction = SceneContraction(order=float("inf"))
 
+        appearance_embedding_dim = self.config.appearance_embed_dim if self.config.use_appearance_embedding else 0
+
         # Fields
         self.field = NerfactoField(
             self.scene_box.aabb,
@@ -163,7 +167,7 @@ class NerfactoModel(Model):
             num_images=self.num_train_data,
             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,
+            appearance_embedding_dim=appearance_embedding_dim,
             average_init_density=self.config.average_init_density,
             implementation=self.config.implementation,
         )