[Datumaro] Optimize mask operations (#1232)

* Optimize mask to rle

* Optimize mask operations

* Fix dm format cmdline

* Use RLE masks in datumaro format

datumaro/datumaro/components/extractor.py

@@ -200,14 +200,12 @@ class Mask(Annotation):
         return self._z_order
 
     def as_class_mask(self, label_id=None):
-        from datumaro.util.mask_tools import make_index_mask
         if label_id is None:
             label_id = self.label
-        return make_index_mask(self.image, label_id)
+        return self.image * label_id
 
     def as_instance_mask(self, instance_id):
-        from datumaro.util.mask_tools import make_index_mask
-        return make_index_mask(self.image, instance_id)
+        return self.image * instance_id
 
     def get_area(self):
         return np.count_nonzero(self.image)

datumaro/datumaro/plugins/coco_format/converter.py

@@ -278,7 +278,10 @@ class _InstancesConverter(_TaskConverter):
 
         is_crowd = mask is not None
         if is_crowd:
-            segmentation = mask
+            segmentation = {
+                'counts': list(int(c) for c in mask['counts']),
+                'size': list(int(c) for c in mask['size'])
+            }
         else:
             segmentation = [list(map(float, p)) for p in polygons]
 

datumaro/datumaro/plugins/datumaro_format/converter.py

@@ -6,16 +6,18 @@
 # pylint: disable=no-self-use
 
 import json
+import numpy as np
 import os
 import os.path as osp
 
 from datumaro.components.converter import Converter
 from datumaro.components.extractor import (
     DEFAULT_SUBSET_NAME, Annotation,
-    Label, Mask, Points, Polygon, PolyLine, Bbox, Caption,
+    Label, Mask, RleMask, Points, Polygon, PolyLine, Bbox, Caption,
     LabelCategories, MaskCategories, PointsCategories
 )
 from datumaro.util.image import save_image
+import pycocotools.mask as mask_utils
 from datumaro.components.cli_plugin import CliPlugin
 
 from .format import DatumaroPath
@@ -40,8 +42,6 @@ class _SubsetWriter:
             'items': [],
         }
 
-        self._next_mask_id = 1
-
     @property
     def categories(self):
         return self._data['categories']
@@ -123,33 +123,22 @@ class _SubsetWriter:
         })
         return converted
 
-    def _save_mask(self, mask):
-        mask_id = None
-        if mask is None:
-            return mask_id
-
-        mask_id = self._next_mask_id
-        self._next_mask_id += 1
-
-        filename = '%d%s' % (mask_id, DatumaroPath.MASK_EXT)
-        masks_dir = osp.join(self._context._annotations_dir,
-            DatumaroPath.MASKS_DIR)
-        os.makedirs(masks_dir, exist_ok=True)
-        path = osp.join(masks_dir, filename)
-        save_image(path, mask)
-        return mask_id
-
     def _convert_mask_object(self, obj):
         converted = self._convert_annotation(obj)
 
-        mask = obj.image
-        mask_id = None
-        if mask is not None:
-            mask_id = self._save_mask(mask)
+        if isinstance(obj, RleMask):
+            rle = obj.rle
+        else:
+            rle = mask_utils.encode(
+                np.require(obj.image, dtype=np.uint8, requirements='F'))
 
         converted.update({
             'label_id': _cast(obj.label, int),
-            'mask_id': _cast(mask_id, int),
+            'rle': {
+                # serialize as compressed COCO mask
+                'counts': rle['counts'].decode('ascii'),
+                'size': list(int(c) for c in rle['size']),
+            }
         })
         return converted
 
@@ -289,6 +278,7 @@ class _Converter:
 class DatumaroConverter(Converter, CliPlugin):
     @classmethod
     def build_cmdline_parser(cls, **kwargs):
+        parser = super().build_cmdline_parser(**kwargs)
         parser.add_argument('--save-images', action='store_true',
             help="Save images (default: %(default)s)")
         return parser

datumaro/datumaro/plugins/datumaro_format/extractor.py

@@ -4,16 +4,14 @@
 # SPDX-License-Identifier: MIT
 
 import json
-import logging as log
 import os.path as osp
 
 from datumaro.components.extractor import (SourceExtractor,
     DEFAULT_SUBSET_NAME, DatasetItem,
-    AnnotationType, Label, Mask, Points, Polygon, PolyLine, Bbox, Caption,
+    AnnotationType, Label, RleMask, Points, Polygon, PolyLine, Bbox, Caption,
     LabelCategories, MaskCategories, PointsCategories
 )
 from datumaro.util.image import Image
-from datumaro.util.mask_tools import lazy_mask
 
 from .format import DatumaroPath
 
@@ -127,19 +125,9 @@ class DatumaroExtractor(SourceExtractor):
 
             elif ann_type == AnnotationType.mask:
                 label_id = ann.get('label_id')
-                mask_id = str(ann.get('mask_id'))
-
-                mask_path = osp.join(self._path, DatumaroPath.ANNOTATIONS_DIR,
-                    DatumaroPath.MASKS_DIR, mask_id + DatumaroPath.MASK_EXT)
-                mask = None
-
-                if osp.isfile(mask_path):
-                    mask = lazy_mask(mask_path)
-                else:
-                    log.warn("Not found mask image file '%s', skipped." % \
-                        mask_path)
-
-                loaded.append(Mask(label=label_id, image=mask,
+                rle = ann['rle']
+                rle['counts'] = rle['counts'].encode('ascii')
+                loaded.append(RleMask(rle=rle, label=label_id,
                     id=ann_id, attributes=attributes, group=group))
 
             elif ann_type == AnnotationType.polyline:

datumaro/datumaro/util/mask_tools.py

@@ -111,15 +111,20 @@ def load_mask(path, inverse_colormap=None):
 def lazy_mask(path, inverse_colormap=None):
     return lazy_image(path, lambda path: load_mask(path, inverse_colormap))
 
-
 def mask_to_rle(binary_mask):
-    counts = []
-    for i, (value, elements) in enumerate(
-            groupby(binary_mask.ravel(order='F'))):
-        # decoding starts from 0
-        if i == 0 and value == 1:
-            counts.append(0)
-        counts.append(len(list(elements)))
+    # walk in row-major order as COCO format specifies
+    bounded = binary_mask.ravel(order='F')
+
+    # add borders to sequence
+    # find boundary positions for sequences and compute their lengths
+    difs = np.diff(bounded, prepend=[1 - bounded[0]], append=[1 - bounded[-1]])
+    counts, = np.where(difs != 0)
+
+    # start RLE encoding from 0 as COCO format specifies
+    if bounded[0] != 0:
+        counts = np.diff(counts, prepend=[0])
+    else:
+        counts = np.diff(counts)
 
     return {
         'counts': counts,
@@ -267,7 +272,7 @@ def find_mask_bbox(mask):
 
 def merge_masks(masks):
     """
-        Merges masks into one, mask order is resposible for z order.
+        Merges masks into one, mask order is responsible for z order.
     """
     if not masks:
         return None

datumaro/tests/test_masks.py

@@ -68,15 +68,7 @@ class PolygonConversionsTest(TestCase):
             self.assertTrue(np.array_equal(e_mask, c_mask),
                 '#%s: %s\n%s\n' % (i, e_mask, c_mask))
 
-    def test_mask_to_rle(self):
-        source_mask = np.array([
-            [0, 1, 1, 1, 0, 1, 1, 1, 1, 0],
-            [0, 0, 1, 1, 0, 1, 0, 1, 0, 0],
-            [0, 0, 0, 1, 0, 1, 1, 0, 0, 0],
-            [0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
-            [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
-        ])
-
+    def _test_mask_to_rle(self, source_mask):
         rle_uncompressed = mask_tools.mask_to_rle(source_mask)
 
         from pycocotools import mask as mask_utils
@@ -87,6 +79,43 @@ class PolygonConversionsTest(TestCase):
         self.assertTrue(np.array_equal(source_mask, resulting_mask),
             '%s\n%s\n' % (source_mask, resulting_mask))
 
+    def test_mask_to_rle_multi(self):
+        cases = [
+            np.array([
+                [0, 1, 1, 1, 0, 1, 1, 1, 1, 0],
+                [0, 0, 1, 1, 0, 1, 0, 1, 0, 0],
+                [0, 0, 0, 1, 0, 1, 1, 0, 0, 0],
+                [0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
+                [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+            ]),
+
+            np.array([
+                [0]
+            ]),
+            np.array([
+                [1]
+            ]),
+
+            np.array([
+                [1, 0, 0, 0, 0, 0, 0, 1, 0, 0],
+                [0, 0, 1, 1, 1, 0, 0, 0, 0, 0],
+                [1, 0, 1, 0, 1, 1, 1, 0, 0, 0],
+                [1, 1, 0, 1, 0, 1, 1, 1, 1, 0],
+                [1, 0, 1, 0, 1, 0, 0, 0, 0, 0],
+                [1, 0, 0, 1, 0, 0, 0, 1, 0, 1],
+                [1, 1, 0, 0, 1, 1, 0, 0, 0, 1],
+                [0, 0, 1, 0, 0, 0, 1, 1, 1, 1],
+                [1, 1, 0, 0, 0, 0, 0, 1, 0, 0],
+                [1, 1, 1, 1, 1, 0, 1, 0, 1, 0],
+                [0, 1, 0, 1, 1, 1, 1, 1, 0, 0],
+                [0, 1, 0, 0, 0, 1, 0, 0, 1, 0],
+                [1, 1, 0, 1, 0, 0, 1, 1, 1, 1],
+            ])
+        ]
+
+        for case in cases:
+            self._test_mask_to_rle(case)
+
 class ColormapOperationsTest(TestCase):
     def test_can_paint_mask(self):
         mask = np.zeros((1, 3), dtype=np.uint8)