Replace mask format support with Datumaro (#1163)

* Add box to mask transform

* Fix 'source' labelmap mode in voc converter

* Import groups

* Replace mask format support

* Update mask format documentation

* codacy

* Fix tests

* Fix dataset

* Fix segments grouping

* Merge instances in mask export

README.md

@@ -38,18 +38,17 @@ Format selection is possible after clicking on the Upload annotation / Dump anno
 [Datumaro](datumaro/README.md) dataset framework allows additional dataset transformations
 via its command line tool.
 
-| Annotation format                                                                  | Dumper | Loader |
-| ---------------------------------------------------------------------------------- | ------ | ------ |
-| [CVAT XML v1.1 for images](cvat/apps/documentation/xml_format.md#annotation)       | X      | X      |
-| [CVAT XML v1.1 for a video](cvat/apps/documentation/xml_format.md#interpolation)   | X      | X      |
-| [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/)                              | X      | X      |
-| [YOLO](https://pjreddie.com/darknet/yolo/)                                         | X      | X      |
-| [MS COCO Object Detection](http://cocodataset.org/#format-data)                    | X      | X      |
-| PNG mask                                                                           | X      |        |
-| PNG instance mask                                                                  | X      |        |
-| [TFrecord](https://www.tensorflow.org/tutorials/load_data/tf_records)              | X      | X      |
-| [MOT](https://motchallenge.net/)                                                   | X      | X      |
-| [LabelMe](http://labelme.csail.mit.edu/Release3.0)                                 | X      | X      |
+| Annotation format                                                                          | Dumper | Loader |
+| ------------------------------------------------------------------------------------------ | ------ | ------ |
+| [CVAT XML v1.1 for images](cvat/apps/documentation/xml_format.md#annotation)               | X      | X      |
+| [CVAT XML v1.1 for a video](cvat/apps/documentation/xml_format.md#interpolation)           | X      | X      |
+| [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/)                                      | X      | X      |
+| [YOLO](https://pjreddie.com/darknet/yolo/)                                                 | X      | X      |
+| [MS COCO Object Detection](http://cocodataset.org/#format-data)                            | X      | X      |
+| PNG class mask + instance mask as in [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/) | X      | X      |
+| [TFrecord](https://www.tensorflow.org/tutorials/load_data/tf_records)                      | X      | X      |
+| [MOT](https://motchallenge.net/)                                                           | X      | X      |
+| [LabelMe](http://labelme.csail.mit.edu/Release3.0)                                         | X      | X      |
 
 ## Links
 - [Intel AI blog: New Computer Vision Tool Accelerates Annotation of Digital Images and Video](https://www.intel.ai/introducing-cvat)

cvat/apps/annotation/README.md

@@ -506,18 +506,48 @@ python create_pascal_tf_record.py --data_dir <path to VOCdevkit> --set train --y
 - downloaded file: a zip archive with the following structure:
   ```bash
   taskname.zip
-  ├── frame_000001.png
-  ├── frame_000002.png
-  ├── frame_000003.png
-  ├── ...
-  └── colormap.txt
+  ├── labelmap.txt # optional, required for non-VOC labels
+  ├── ImageSets/
+  │   └── Segmentation/
+  │       └── default.txt # list of image names without extension
+  ├── SegmentationClass/ # merged class masks
+  │   └── image1.png
+  │   └── image2.png
+  └── SegmentationObject/ # merged instance masks
+      └── image1.png
+      └── image2.png
   ```
   Mask is a png image with several (RGB) channels where each pixel has own color which corresponds to a label.
   Color generation correspond to the Pascal VOC color generation
   [algorithm](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/htmldoc/devkit_doc.html#sec:voclabelcolormap).
   (0, 0, 0) is used for background.
-  `colormap.txt` file contains the values of the used colors in RGB format.
+  `labelmap.txt` file contains the values of the used colors in RGB format. The file structure:
+  ```bash
+  # label:color_rgb:parts:actions
+  background:0,128,0::
+  aeroplane:10,10,128::
+  bicycle:10,128,0::
+  bird:0,108,128::
+  boat:108,0,100::
+  bottle:18,0,8::
+  bus:12,28,0::
+  ```
 - supported shapes - Rectangles, Polygons
 
 #### Mask loader description
-Not supported
+- uploaded file: a zip archive of the following structure:
+  ```bash
+  name.zip
+  ├── labelmap.txt # optional, required for non-VOC labels
+  ├── ImageSets/
+  │   └── Segmentation/
+  │       └── <any_subset_name>.txt
+  ├── SegmentationClass/
+  │   └── image1.png
+  │   └── image2.png
+  └── SegmentationObject/
+      └── image.png
+      └── image2.png
+  ```
+- supported shapes: Polygons
+- additional comments: the CVAT task should be created with the full label set that may be in the annotation files

cvat/apps/annotation/mask.py

@@ -6,130 +6,60 @@
     "name": "MASK",
     "dumpers": [
         {
-            "display_name": "{name} (by class) {format} {version}",
+            "display_name": "{name} {format} {version}",
             "format": "ZIP",
-            "version": "1.0",
-            "handler": "dump_by_class"
+            "version": "1.1",
+            "handler": "dump",
         },
+    ],
+    "loaders": [
         {
-            "display_name": "{name} (by instance) {format} {version}",
+            "display_name": "{name} {format} {version}",
             "format": "ZIP",
-            "version": "1.0",
-            "handler": "dump_by_instance"
+            "version": "1.1",
+            "handler": "load",
         },
     ],
-    "loaders": [
-    ],
 }
 
-MASK_BY_CLASS = 0
-MASK_BY_INSTANCE = 1
-
-def convert_box_to_polygon(shape):
-        xtl = shape.points[0]
-        ytl = shape.points[1]
-        xbr = shape.points[2]
-        ybr = shape.points[3]
-
-        return [xtl, ytl, xbr, ytl, xbr, ybr, xtl, ybr]
-
-def create_mask_colorizer(annotations, colorize_type):
-    import numpy as np
-    from collections import OrderedDict
-
-    class MaskColorizer:
-
-        def __init__(self, annotations, colorize_type):
-
-            if colorize_type == MASK_BY_CLASS:
-                self.colors = self.gen_class_mask_colors(annotations)
-            elif colorize_type == MASK_BY_INSTANCE:
-                self.colors = self.gen_instance_mask_colors()
-
-        def generate_pascal_colormap(self, size=256):
-            # RGB format, (0, 0, 0) used for background
-            colormap = np.zeros((size, 3), dtype=int)
-            ind = np.arange(size, dtype=int)
-
-            for shift in reversed(range(8)):
-                for channel in range(3):
-                    colormap[:, channel] |= ((ind >> channel) & 1) << shift
-                ind >>= 3
-
-            return colormap
-
-        def gen_class_mask_colors(self, annotations):
-            colormap = self.generate_pascal_colormap()
-            labels = [label[1]["name"] for label in annotations.meta["task"]["labels"] if label[1]["name"] != 'background']
-            labels.insert(0, 'background')
-            label_colors = OrderedDict((label, colormap[idx]) for idx, label in enumerate(labels))
-
-            return label_colors
-
-        def gen_instance_mask_colors(self):
-            colormap = self.generate_pascal_colormap()
-            # The first color is black
-            instance_colors = OrderedDict((idx, colormap[idx]) for idx in range(len(colormap)))
-
-            return instance_colors
-
-    return MaskColorizer(annotations, colorize_type)
-
-def dump(file_object, annotations, colorize_type):
-
-    from zipfile import ZipFile, ZIP_STORED
-    import numpy as np
-    import os
-    from pycocotools import mask as maskUtils
-    import matplotlib.image
-    import io
-
-    colorizer = create_mask_colorizer(annotations, colorize_type=colorize_type)
-    if colorize_type == MASK_BY_CLASS:
-        save_dir = "SegmentationClass"
-    elif colorize_type == MASK_BY_INSTANCE:
-        save_dir = "SegmentationObject"
-
-    with ZipFile(file_object, "w", ZIP_STORED) as output_zip:
-        for frame_annotation in annotations.group_by_frame():
-            image_name = frame_annotation.name
-            annotation_name = "{}.png".format(os.path.splitext(os.path.basename(image_name))[0])
-            width = frame_annotation.width
-            height = frame_annotation.height
-
-            shapes = frame_annotation.labeled_shapes
-            # convert to mask only rectangles and polygons
-            shapes = [shape for shape in shapes if shape.type == 'rectangle' or shape.type == 'polygon']
-            if not shapes:
-                continue
-            shapes = sorted(shapes, key=lambda x: int(x.z_order))
-            img_mask = np.zeros((height, width, 3))
-            buf_mask = io.BytesIO()
-            for shape_index, shape in enumerate(shapes):
-                points = shape.points if shape.type != 'rectangle' else convert_box_to_polygon(shape)
-                rles = maskUtils.frPyObjects([points], height, width)
-                rle = maskUtils.merge(rles)
-                mask = maskUtils.decode(rle)
-                idx = (mask > 0)
-                # get corresponding color
-                if colorize_type == MASK_BY_CLASS:
-                    color = colorizer.colors[shape.label] / 255
-                elif colorize_type == MASK_BY_INSTANCE:
-                    color = colorizer.colors[shape_index+1] / 255
-
-                img_mask[idx] = color
-
-            # write mask
-            matplotlib.image.imsave(buf_mask, img_mask, format='png')
-            output_zip.writestr(os.path.join(save_dir, annotation_name), buf_mask.getvalue())
-        # Store color map for each class
-        labels = '\n'.join('{}:{}'.format(label, ','.join(str(i) for i in color)) for label, color in colorizer.colors.items())
-        output_zip.writestr('colormap.txt', labels)
-
-def dump_by_class(file_object, annotations):
-
-    return dump(file_object, annotations, MASK_BY_CLASS)
-
-def dump_by_instance(file_object, annotations):
 
-    return dump(file_object, annotations, MASK_BY_INSTANCE)
+def dump(file_object, annotations):
+    from cvat.apps.dataset_manager.bindings import CvatAnnotationsExtractor
+    from cvat.apps.dataset_manager.util import make_zip_archive
+    from datumaro.components.project import Environment, Dataset
+    from tempfile import TemporaryDirectory
+
+    env = Environment()
+    polygons_to_masks = env.transforms.get('polygons_to_masks')
+    boxes_to_masks = env.transforms.get('boxes_to_masks')
+    merge_instance_segments = env.transforms.get('merge_instance_segments')
+    id_from_image = env.transforms.get('id_from_image_name')
+
+    extractor = CvatAnnotationsExtractor('', annotations)
+    extractor = extractor.transform(polygons_to_masks)
+    extractor = extractor.transform(boxes_to_masks)
+    extractor = extractor.transform(merge_instance_segments)
+    extractor = extractor.transform(id_from_image)
+    extractor = Dataset.from_extractors(extractor) # apply lazy transforms
+    converter = env.make_converter('voc_segmentation',
+        apply_colormap=True, label_map='source')
+    with TemporaryDirectory() as temp_dir:
+        converter(extractor, save_dir=temp_dir)
+        make_zip_archive(temp_dir, file_object)
+
+def load(file_object, annotations):
+    from pyunpack import Archive
+    from tempfile import TemporaryDirectory
+    from datumaro.plugins.voc_format.importer import VocImporter
+    from datumaro.components.project import Environment
+    from cvat.apps.dataset_manager.bindings import import_dm_annotations
+
+    archive_file = file_object if isinstance(file_object, str) else getattr(file_object, "name")
+    with TemporaryDirectory() as tmp_dir:
+        Archive(archive_file).extractall(tmp_dir)
+
+        dm_project = VocImporter()(tmp_dir)
+        dm_dataset = dm_project.make_dataset()
+        masks_to_polygons = Environment().transforms.get('masks_to_polygons')
+        dm_dataset = dm_dataset.transform(masks_to_polygons)
+        import_dm_annotations(dm_dataset, annotations)

cvat/apps/dataset_manager/bindings.py

@@ -211,6 +211,22 @@ def import_dm_annotations(dm_dataset, cvat_task_anno):
     for item in dm_dataset:
         frame_number = match_frame(item, cvat_task_anno)
 
+        # do not store one-item groups
+        group_map = { 0: 0 }
+        group_size = { 0: 0 }
+        for ann in item.annotations:
+            if ann.type in shapes:
+                group = group_map.get(ann.group)
+                if group is None:
+                    group = len(group_map)
+                    group_map[ann.group] = group
+                    group_size[ann.group] = 1
+                else:
+                    group_size[ann.group] += 1
+        group_map = {g: s for g, s in group_size.items()
+            if 1 < s and group_map[g]}
+        group_map = {g: i for i, g in enumerate([0] + sorted(group_map))}
+
         for ann in item.annotations:
             if ann.type in shapes:
                 cvat_task_anno.add_shape(cvat_task_anno.LabeledShape(
@@ -219,5 +235,6 @@ def import_dm_annotations(dm_dataset, cvat_task_anno):
                     label=label_cat.items[ann.label].name,
                     points=ann.points,
                     occluded=False,
+                    group=group_map.get(ann.group, 0),
                     attributes=[],
                 ))

cvat/apps/engine/tests/test_rest_api.py

@@ -2658,9 +2658,9 @@ def _get_initial_annotation(annotation_format):
             elif annotation_format == "COCO JSON 1.0":
                 annotations["shapes"] = polygon_shapes_wo_attrs
 
-            elif annotation_format == "MASK ZIP 1.0":
-                annotations["shapes"] = rectangle_shapes_with_attrs + rectangle_shapes_wo_attrs + polygon_shapes_wo_attrs
-                annotations["tracks"] = rectangle_tracks_with_attrs + rectangle_tracks_wo_attrs
+            elif annotation_format == "MASK ZIP 1.1":
+                annotations["shapes"] = rectangle_shapes_wo_attrs + polygon_shapes_wo_attrs
+                annotations["tracks"] = rectangle_tracks_wo_attrs
 
             elif annotation_format == "MOT CSV 1.0":
                 annotations["tracks"] = rectangle_tracks_wo_attrs
@@ -2730,6 +2730,8 @@ def _get_initial_annotation(annotation_format):
                     }
 
                     for loader in annotation_format["loaders"]:
+                        if loader["display_name"] == "MASK ZIP 1.1":
+                            continue # can't really predict the result and check
                         response = self._upload_api_v1_tasks_id_annotations(task["id"], annotator, uploaded_data, "format={}".format(loader["display_name"]))
                         self.assertEqual(response.status_code, HTTP_202_ACCEPTED)
 

datumaro/datumaro/components/project.py

@@ -321,17 +321,17 @@ def from_extractors(cls, *sources):
         subsets = defaultdict(lambda: Subset(dataset))
         for source in sources:
             for item in source:
-                path = None # NOTE: merge everything into our own dataset
-
                 existing_item = subsets[item.subset].items.get(item.id)
                 if existing_item is not None:
-                    item = self._merge_items(existing_item, item, path=path)
-                else:
-                    item = item.wrap(path=path, annotations=item.annotations)
+                    path = existing_item.path
+                    if item.path != path:
+                        path = None
+                    item = cls._merge_items(existing_item, item, path=path)
 
                 subsets[item.subset].items[item.id] = item
 
-        self._subsets = dict(subsets)
+        dataset._subsets = dict(subsets)
+        return dataset
 
     def __init__(self, categories=None):
         super().__init__()
@@ -419,7 +419,7 @@ def _merge_items(cls, existing_item, current_item, path=None):
                     image._path = current_item.image.path
 
             if all([existing_item.image._size, current_item.image._size]):
-                assert existing_item.image._size == current_item.image._size, "Image info differs for item '%s'" % item.id
+                assert existing_item.image._size == current_item.image._size, "Image info differs for item '%s'" % existing_item.id
             elif existing_item.image._size:
                 image._size = existing_item.image._size
             else:

datumaro/datumaro/plugins/coco_format/converter.py

@@ -361,6 +361,7 @@ def save_annotations(self, item):
 
     @classmethod
     def find_solitary_points(cls, annotations):
+        annotations = sorted(annotations, key=lambda a: a.group)
         solitary_points = []
 
         for g_id, group in groupby(annotations, lambda a: a.group):

datumaro/datumaro/plugins/transforms.py

@@ -181,6 +181,27 @@ def convert_polygon(polygon, img_h, img_w):
         return RleMask(rle=rle, label=polygon.label, z_order=polygon.z_order,
             id=polygon.id, attributes=polygon.attributes, group=polygon.group)
 
+class BoxesToMasks(Transform, CliPlugin):
+    def transform_item(self, item):
+        annotations = []
+        for ann in item.annotations:
+            if ann.type == AnnotationType.bbox:
+                if not item.has_image:
+                    raise Exception("Image info is required for this transform")
+                h, w = item.image.size
+                annotations.append(self.convert_bbox(ann, h, w))
+            else:
+                annotations.append(ann)
+
+        return self.wrap_item(item, annotations=annotations)
+
+    @staticmethod
+    def convert_bbox(bbox, img_h, img_w):
+        rle = mask_utils.frPyObjects([bbox.as_polygon()], img_h, img_w)[0]
+
+        return RleMask(rle=rle, label=bbox.label, z_order=bbox.z_order,
+            id=bbox.id, attributes=bbox.attributes, group=bbox.group)
+
 class MasksToPolygons(Transform, CliPlugin):
     def transform_item(self, item):
         annotations = []