Adapt cusfm format for training

Author: ApdowJN

configs/apps/cusfm_3dgut_mcmc.yaml

@@ -0,0 +1,15 @@
+# @package _global_
+
+# order in which configs override each other (/* - denotes a relative search path)
+defaults:
+  - /base_mcmc
+  - /dataset: colmap
+  - /initialization: accumulated_point_cloud
+  - /render: 3dgut
+  - _self_
+
+# overwrite of default parameters
+val_frequency: 999999 # never validate
+
+initialization:
+  accumulated_point_cloud_path: ???  # Path to accumulated point cloud PLY file to be provided by user at runtime

configs/initialization/accumulated_point_cloud.yaml

@@ -0,0 +1,5 @@
+method: accumulated_point_cloud
+observation_scale_factor: 0.01
+use_observation_points: true
+accumulated_point_cloud_path: ???  # Path to accumulated point cloud PLY file
+

threedgrut/datasets/utils.py

@@ -479,7 +479,8 @@ def read_colmap_extrinsics_text(path):
     with open(path, "r") as fid:
         # Skip comment lines and get valid lines
         lines = (line.strip() for line in fid)
-        lines = (line for line in lines if line and not line.startswith("#"))
+        # lines = (line for line in lines if line and not line.startswith("#"))
+        lines = (line for line in lines if line == "" or not line.startswith("#"))
         # Process lines in pairs (image info + points info)
         try:
             while True:

threedgrut/model/model.py

@@ -252,6 +252,55 @@ def init_from_colmap(self, root_path: str, observer_pts):
             file_pts, observer_pts, file_rgb, use_observer_pts=self.conf.initialization.use_observation_points
         )
 
+    def init_from_accumulated_point_cloud(self, pc_path: str, observer_pts):
+        """
+        Initialize gaussians from an accumulated point cloud PLY file.
+        Similar to init_from_colmap but loads from a given PLY file instead of sparse/0/points3D.txt
+        
+        Args:
+            pc_path: Path to the PLY point cloud file
+            observer_pts: Observer points tensor for scale initialization
+        """
+        logger.info(f"Loading accumulated point cloud from {pc_path}...")
+        
+        # Read PLY file
+        plydata = PlyData.read(pc_path)
+        vertices = plydata['vertex']
+        
+        # Extract XYZ coordinates
+        xyz = np.stack([
+            vertices['x'],
+            vertices['y'],
+            vertices['z']
+        ], axis=1).astype(np.float32)
+        
+        # Extract RGB colors (check if they exist)
+        if 'red' in vertices and 'green' in vertices and 'blue' in vertices:
+            rgb = np.stack([
+                vertices['red'],
+                vertices['green'],
+                vertices['blue']
+            ], axis=1).astype(np.uint8)
+        else:
+            # If no colors, initialize with random colors
+            logger.warning("No RGB data found in point cloud, using random colors")
+            rgb = np.random.randint(0, 256, size=(len(vertices), 3), dtype=np.uint8)
+        
+        # Convert to torch tensors
+        file_pts = torch.tensor(xyz, dtype=torch.float32, device=self.device)
+        file_rgb = torch.tensor(rgb, dtype=torch.uint8, device=self.device)
+        
+        logger.info(f"Loaded {len(file_pts)} points from accumulated point cloud")
+        
+        # Initialize using the same method as COLMAP
+        assert file_rgb.dtype == torch.uint8, "Expecting RGB values to be in [0, 255] range"
+        self.default_initialize_from_points(
+            file_pts, 
+            observer_pts, 
+            file_rgb, 
+            use_observer_pts=self.conf.initialization.use_observation_points
+        )
+        
     def init_from_pretrained_point_cloud(self, pc_path: str, set_optimizable_parameters: bool = True):
         data = PlyData.read(pc_path)
         num_gaussians = len(data["vertex"])

threedgrut/trainer.py

@@ -256,6 +256,13 @@ def setup_training(self, conf: DictConfig, model: MixtureOfGaussians, train_data
                         train_dataset.get_observer_points(), dtype=torch.float32, device=self.device
                     )
                     model.init_from_colmap(conf.path, observer_points)
+                case "accumulated_point_cloud":
+                    observer_points = torch.tensor(
+                        train_dataset.get_observer_points(), dtype=torch.float32, device=self.device
+                    )
+                    ply_path = conf.initialization.accumulated_point_cloud_path
+                    logger.info(f"Initializing from accumulated point cloud: {ply_path}")
+                    model.init_from_accumulated_point_cloud(ply_path, observer_points)
                 case "point_cloud":
                     try:
                         ply_path = os.path.join(conf.path, "point_cloud.ply")