[FIX] Dataset Adapter: Avoid duplicated annotation and permit empty image

otx/core/data/adapter/detection_dataset_adapter.py

@@ -13,6 +13,7 @@
 from otx.api.entities.datasets import DatasetEntity
 from otx.api.entities.image import Image
 from otx.api.entities.model_template import TaskType
+from otx.api.entities.subset import Subset
 from otx.core.data.adapter.base_dataset_adapter import BaseDatasetAdapter
 
 
@@ -48,8 +49,13 @@ def get_otx_dataset(self) -> DatasetEntity:
 
                         if ann.label not in used_labels:
                             used_labels.append(ann.label)
-                    dataset_item = DatasetItemEntity(image, self._get_ann_scene_entity(shapes), subset=subset)
-                    dataset_items.append(dataset_item)
 
+                    if (
+                        len(shapes) > 0
+                        or subset == Subset.UNLABELED
+                        or (subset != Subset.TRAINING and len(datumaro_item.annotations) == 0)
+                    ):
+                        dataset_item = DatasetItemEntity(image, self._get_ann_scene_entity(shapes), subset=subset)
+                        dataset_items.append(dataset_item)
         self.remove_unused_label_entities(used_labels)
         return DatasetEntity(items=dataset_items)

otx/core/data/adapter/segmentation_dataset_adapter.py

@@ -92,6 +92,7 @@ def get_otx_dataset(self) -> DatasetEntity:
                                 shapes.append(self._get_polygon_entity(d_polygon, image.width, image.height))
                                 if d_polygon.label not in used_labels:
                                     used_labels.append(d_polygon.label)
+
                     if len(shapes) > 0 or subset == Subset.UNLABELED:
                         dataset_item = DatasetItemEntity(image, self._get_ann_scene_entity(shapes), subset=subset)
                         dataset_items.append(dataset_item)

tests/assets/car_tree_bug/annotations/instances_val.json

@@ -28,7 +28,17 @@
       "id": 8,
       "width": 1280,
       "height": 720,
-      "file_name": "Slide20.PNG",
+      "file_name": "Slide4.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 9,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide5.PNG",
       "license": 0,
       "flickr_url": "",
       "coco_url": "",

tests/unit/core/data/adapter/test_detection_adapter.py

@@ -3,8 +3,8 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 import os
-from typing import Optional
 
+from otx.api.entities.annotation import NullAnnotationSceneEntity
 from otx.api.entities.datasets import DatasetEntity
 from otx.api.entities.label_schema import LabelSchemaEntity
 from otx.api.entities.model_template import TaskType
@@ -20,42 +20,48 @@
 class TestOTXDetectionDatasetAdapter:
     def setup_method(self):
         self.root_path = os.getcwd()
+
+    @e2e_pytest_unit
+    def test_detection(self):
         task = "detection"
 
-        self.task_type: TaskType = TASK_NAME_TO_TASK_TYPE[task]
+        task_type: TaskType = TASK_NAME_TO_TASK_TYPE[task]
         data_root_dict: dict = TASK_NAME_TO_DATA_ROOT[task]
 
-        self.train_data_roots: str = os.path.join(self.root_path, data_root_dict["train"])
-        self.val_data_roots: str = os.path.join(self.root_path, data_root_dict["val"])
-        self.test_data_roots: str = os.path.join(self.root_path, data_root_dict["test"])
-        self.unlabeled_data_roots: Optional[str] = None
-        if "unlabeled" in data_root_dict:
-            self.unlabeled_data_roots = os.path.join(self.root_path, data_root_dict["unlabeled"])
-
-        self.train_dataset_adapter = DetectionDatasetAdapter(
-            task_type=self.task_type,
-            train_data_roots=self.train_data_roots,
-            val_data_roots=self.val_data_roots,
-            unlabeled_data_roots=self.unlabeled_data_roots,
-        )
+        train_data_roots: str = os.path.join(self.root_path, data_root_dict["train"])
+        val_data_roots: str = os.path.join(self.root_path, data_root_dict["val"])
+        test_data_roots: str = os.path.join(self.root_path, data_root_dict["test"])
 
-        self.test_dataset_adapter = DetectionDatasetAdapter(
-            task_type=self.task_type,
-            test_data_roots=self.test_data_roots,
+        det_train_dataset_adapter = DetectionDatasetAdapter(
+            task_type=task_type,
+            train_data_roots=train_data_roots,
+            val_data_roots=val_data_roots,
         )
 
-    @e2e_pytest_unit
-    def test_init(self):
-        assert Subset.TRAINING in self.train_dataset_adapter.dataset
-        assert Subset.VALIDATION in self.train_dataset_adapter.dataset
-        assert Subset.TESTING in self.test_dataset_adapter.dataset
-        if self.unlabeled_data_roots is not None:
-            assert Subset.UNLABELED in self.train_dataset_adapter.dataset
+        assert Subset.TRAINING in det_train_dataset_adapter.dataset
+        assert Subset.VALIDATION in det_train_dataset_adapter.dataset
 
-    @e2e_pytest_unit
-    def test_get_otx_dataset(self):
-        assert isinstance(self.train_dataset_adapter.get_otx_dataset(), DatasetEntity)
-        assert isinstance(self.test_dataset_adapter.get_otx_dataset(), DatasetEntity)
+        det_train_dataset = det_train_dataset_adapter.get_otx_dataset()
+        det_train_label_schema = det_train_dataset_adapter.get_label_schema()
+        assert isinstance(det_train_dataset, DatasetEntity)
+        assert isinstance(det_train_label_schema, LabelSchemaEntity)
+
+        # In the test data, there is a empty_label image.
+        # So, has_empty_label should be True
+        has_empty_label = False
+        for train_data in det_train_dataset:
+            if isinstance(train_data.annotation_scene, NullAnnotationSceneEntity):
+                has_empty_label = True
+        assert has_empty_label is True
+
+        det_test_dataset_adapter = DetectionDatasetAdapter(
+            task_type=task_type,
+            test_data_roots=test_data_roots,
+        )
+
+        assert Subset.TESTING in det_test_dataset_adapter.dataset
+        assert isinstance(det_test_dataset_adapter.get_otx_dataset(), DatasetEntity)
+        assert isinstance(det_test_dataset_adapter.get_label_schema(), LabelSchemaEntity)
 
     @e2e_pytest_unit
     def test_instance_segmentation(self):
@@ -77,8 +83,18 @@ def test_instance_segmentation(self):
         assert Subset.TRAINING in instance_seg_train_dataset_adapter.dataset
         assert Subset.VALIDATION in instance_seg_train_dataset_adapter.dataset
 
-        assert isinstance(instance_seg_train_dataset_adapter.get_otx_dataset(), DatasetEntity)
-        assert isinstance(instance_seg_train_dataset_adapter.get_label_schema(), LabelSchemaEntity)
+        instance_seg_otx_train_data = instance_seg_train_dataset_adapter.get_otx_dataset()
+        instance_seg_otx_train_label_schema = instance_seg_train_dataset_adapter.get_label_schema()
+        assert isinstance(instance_seg_otx_train_data, DatasetEntity)
+        assert isinstance(instance_seg_otx_train_label_schema, LabelSchemaEntity)
+
+        # In the test data, there is a empty_label image.
+        # So, has_empty_label should be True
+        has_empty_label = False
+        for train_data in instance_seg_otx_train_data:
+            if isinstance(train_data.annotation_scene, NullAnnotationSceneEntity):
+                has_empty_label = True
+        assert has_empty_label is True
 
         instance_seg_test_dataset_adapter = DetectionDatasetAdapter(
             task_type=task_type,