in a process of updating seans model to work with 4 or 6 classes

run_bash_scripts/sem_seg_sean.sh

@@ -3,12 +3,12 @@
 ############################ parameters #################################################
 # General parameters
 CLEAR_INPUT_FOLDER=1  # 1: clear input folder, 0: not clear input folder
-CONDA_ENV="pdal-env" # conda environment for running the pipeline
+CONDA_ENV="pdal-env-1" # conda environment for running the pipeline
 
 # Parameters for the semetnic segmentation
 data_folder="" # path to the folder containing the data
-checkpoint_model_path="./fsct/model/model.pth"
-batch_size=5 # batch size for the inference
+checkpoint_model_path="./fsct/model/model.pth" # path to the checkpoint model (take our basic model as default)
+batch_size=10 # batch size for the inference
 tile_size=10 # tile size in meters
 min_density=75 # minimum density of points in a tile(used for removing small tiles)
 remove_small_tiles=0 # 1: remove small tiles, 0: not remove small tiles
@@ -45,6 +45,11 @@ echo "d: data_folder"
 echo "      The values of the parameters:"
 echo "data_folder: $data_folder"
 echo "remove_small_tiles: $remove_small_tiles"
+echo "checkpoint_model_path: $checkpoint_model_path"
+echo "batch_size: $batch_size"
+echo "tile_size: $tile_size"
+echo "min_density: $min_density"
+
 
 # Do the environment setup
 # check if PYTHONPATH is set to the current directory

sean_sem_seg/inference.py

@@ -19,6 +19,10 @@ from tools import load_file, save_file
 import shutil
 import sys
 
+from other_parameters import other_parameters
+
+NUM_CLASSES = other_parameters['num_classes']
+
 sys.setrecursionlimit(10**8)  # Can be necessary for dealing with large point clouds.
 
 
@@ -60,11 +64,12 @@ def choose_most_confident_label(point_cloud, original_point_cloud):
     )
     _, indices = neighbours.kneighbors(original_point_cloud[:, :3])
 
-    labels = np.zeros((original_point_cloud.shape[0], 5))
-    labels[:, :4] = np.median(point_cloud[indices][:, :, -4:], axis=1)
-    labels[:, 4] = np.argmax(labels[:, :4], axis=1)
+    labels = np.zeros((original_point_cloud.shape[0], NUM_CLASSES + 1))
+    labels[:, :NUM_CLASSES] = np.median(point_cloud[indices][:, :, -NUM_CLASSES:], axis=1)
+    labels[:, NUM_CLASSES] = np.argmax(labels[:, :NUM_CLASSES], axis=1)
 
     original_point_cloud = np.hstack((original_point_cloud, labels[:, 4:]))
+
     return original_point_cloud
 
 class SemanticSegmentation:
@@ -99,7 +104,7 @@ class SemanticSegmentation:
 
         test_loader = DataLoader(test_dataset, batch_size=self.parameters["batch_size"], shuffle=False, num_workers=0)
 
-        model = Net(num_classes=4).to(self.device)
+        model = Net(num_classes=NUM_CLASSES).to(self.device)
         if self.parameters["use_CPU_only"]:
             model.load_state_dict(
                 torch.load(

sean_sem_seg/other_parameters.py

@@ -9,6 +9,7 @@ other_parameters = dict(
     box_overlap=[0.5, 0.5, 0.5],  # Overlap of the sliding box used for semantic segmentation.
     min_points_per_box=1000,  # Minimum number of points for input to the model. Too few points and it becomes near impossible to accurately label them (though assuming vegetation class is the safest bet here).
     max_points_per_box=20000,  # Maximum number of points for input to the model. The model may tolerate higher numbers if you decrease the batch size accordingly (to fit on the GPU), but this is not tested.
+    num_classes = 4,  # Number of classes in the model. Don't change this unless you are changing the model.
     noise_class=0,  # Don't change
     terrain_class=1,  # Don't change
     vegetation_class=2,  # Don't change

sean_sem_seg/post_segmentation_script.py

@@ -106,6 +106,8 @@ class PostProcessing:
         return grid_points
 
     def process_point_cloud(self):
+        print("Processing point cloud...")
+
         self.terrain_points = self.point_cloud[
             self.point_cloud[:, self.label_index] == self.terrain_class_label
         ]  # -2 is now the class label as we added the height above DTM column.