[Datumaro] Change alignment in mask parsing (#1547)

CHANGELOG.md

@@ -40,6 +40,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Task/Job buttons has no "Open in new tab" option (<https://github.com/opencv/cvat/pull/1419>)
 - Delete point context menu option has no shortcut hint (<https://github.com/opencv/cvat/pull/1416>)
 - Fixed issue with unnecessary tag activation in cvat-canvas (<https://github.com/opencv/cvat/issues/1540>)
+- Fixed an issue with large number of instances in instance mask (https://github.com/opencv/cvat/issues/1539)
 - Fixed full COCO dataset import error with conflicting labels in keypoints and detection (https://github.com/opencv/cvat/pull/1548)
 - Fixed COCO keypoints skeleton parsing and saving (https://github.com/opencv/cvat/issues/1539)
 

datumaro/datumaro/components/extractor.py

@@ -270,7 +270,7 @@ class CompiledMask:
         if instance_ids is not None:
             assert len(instance_ids) == len(instance_masks)
         else:
-            instance_ids = [1 + i for i in range(len(instance_masks))]
+            instance_ids = range(1, len(instance_masks) + 1)
 
         if instance_labels is not None:
             assert len(instance_labels) == len(instance_masks)
@@ -310,15 +310,13 @@ class CompiledMask:
     def instance_count(self):
         return int(self.instance_mask.max())
 
-    def get_instance_labels(self, class_count=None):
+    def get_instance_labels(self):
-        if class_count is None:
+        class_shift = 16
-            class_count = np.max(self.class_mask) + 1
+        m = (self.class_mask.astype(np.uint32) << class_shift) \
-
+            + self.instance_mask.astype(np.uint32)
-        m = self.class_mask * class_count + self.instance_mask
-        m = m.astype(int)
         keys = np.unique(m)
-        instance_labels = {k % class_count: k // class_count
+        instance_labels = {k & ((1 << class_shift) - 1): k >> class_shift
-            for k in keys if k % class_count != 0
+            for k in keys if k & ((1 << class_shift) - 1) != 0
         }
         return instance_labels
 

datumaro/datumaro/plugins/voc_format/extractor.py

@@ -251,8 +251,7 @@ class VocSegmentationExtractor(_VocExtractor):
 
             if class_mask is not None:
                 label_cat = self._categories[AnnotationType.label]
-                instance_labels = compiled_mask.get_instance_labels(
+                instance_labels = compiled_mask.get_instance_labels()
-                    class_count=len(label_cat.items))
             else:
                 instance_labels = {i: None
                     for i in range(compiled_mask.instance_count)}

datumaro/tests/test_masks.py

@@ -3,6 +3,7 @@ import numpy as np
 from unittest import TestCase
 
 import datumaro.util.mask_tools as mask_tools
+from datumaro.components.extractor import CompiledMask
 
 
 class PolygonConversionsTest(TestCase):
@@ -184,3 +185,13 @@ class ColormapOperationsTest(TestCase):
 
         self.assertTrue(np.array_equal(expected, actual),
             '%s\nvs.\n%s' % (expected, actual))
+
+    def test_can_decode_compiled_mask(self):
+        class_idx = 1000
+        instance_idx = 10000
+        mask = np.array([1])
+        compiled_mask = CompiledMask(mask * class_idx, mask * instance_idx)
+
+        labels = compiled_mask.get_instance_labels()
+
+        self.assertEqual({instance_idx: class_idx}, labels)

datumaro/tests/test_voc_format.py

@@ -444,6 +444,41 @@ class VocConverterTest(TestCase):
                 VocSegmentationConverter(label_map='voc'), test_dir,
                 target_dataset=DstExtractor())
 
+    def test_can_save_voc_segm_with_many_instances(self):
+        def bit(x, y, shape):
+            mask = np.zeros(shape)
+            mask[y, x] = 1
+            return mask
+
+        class TestExtractor(TestExtractorBase):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id=1, subset='a', annotations=[
+                        Mask(image=bit(x, y, shape=[10, 10]),
+                            label=self._label(VOC.VocLabel(3).name),
+                            z_order=10 * y + x + 1
+                        )
+                        for y in range(10) for x in range(10)
+                    ]),
+                ])
+
+        class DstExtractor(TestExtractorBase):
+            def __iter__(self):
+                return iter([
+                    DatasetItem(id=1, subset='a', annotations=[
+                        Mask(image=bit(x, y, shape=[10, 10]),
+                            label=self._label(VOC.VocLabel(3).name),
+                            group=10 * y + x + 1
+                        )
+                        for y in range(10) for x in range(10)
+                    ]),
+                ])
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(TestExtractor(),
+                VocSegmentationConverter(label_map='voc'), test_dir,
+                target_dataset=DstExtractor())
+
     def test_can_save_voc_layout(self):
         class TestExtractor(TestExtractorBase):
             def __iter__(self):