[FIX] Fix label mismatch of evaluation and validation with large dataset in semantic segmentation

otx/algorithms/segmentation/adapters/mmseg/data/dataset.py

@@ -127,7 +127,7 @@ def __init__(
         self.label_map = None
 
         dataset_labels = self.otx_dataset.get_labels(include_empty=False)
-        self.project_labels = self.filter_labels(dataset_labels, classes)
+        self.project_labels = sorted(self.filter_labels(dataset_labels, classes))
         self.CLASSES, self.PALETTE = self.get_classes_and_palette(classes, None)
 
         # Instead of using list data_infos as in CustomDataset, this implementation of dataset
@@ -144,8 +144,8 @@ def __init__(
 
     @staticmethod
     @check_input_parameters_type()
-    def filter_labels(all_labels: List[LabelEntity], label_names: List[str]):
-        """Filtering Labels function."""
+    def filter_labels(all_labels: List[LabelEntity], label_names: List[str]) -> List[LabelEntity]:
+        """Filter and collect actual label entities."""
         filtered_labels = []
         for label_name in label_names:
             matches = [label for label in all_labels if label.name == label_name]
@@ -263,6 +263,10 @@ def __init__(self, **kwargs):
         else:
             classes = []
         super().__init__(otx_dataset=otx_dataset, pipeline=pipeline, classes=classes)
+
+        # TODO [Soobee] : Need to align again when background label is removed
+        self.CLASSES = ["background"] + [label.name for label in self.project_labels]
+
         if self.label_map is None:
             self.label_map = {}
             for i, c in enumerate(self.CLASSES):

otx/algorithms/segmentation/tasks/inference.py

@@ -93,7 +93,7 @@ def __init__(self, task_environment: TaskEnvironment, **kwargs):
         self._label_dictionary = {}  # type: Dict
 
         super().__init__(SegmentationConfig, task_environment, **kwargs)
-        self._label_dictionary = dict(enumerate(self._labels, 1))
+        self._label_dictionary = dict(enumerate(sorted(self._labels), 1))
 
     @check_input_parameters_type({"dataset": DatasetParamTypeCheck})
     def infer(

otx/api/usecases/evaluation/basic_operations.py

@@ -34,6 +34,8 @@ def get_intersections_and_cardinalities(
     Returns:
         Tuple[NumberPerLabel, NumberPerLabel]: (all_intersections, all_cardinalities)
     """
+
+    # TODO [Soobee] : Add score for background label and align the calculation method with validation
     all_intersections: NumberPerLabel = {label: 0 for label in labels}
     all_intersections[None] = 0
     all_cardinalities: NumberPerLabel = {label: 0 for label in labels}