[Datumaro] COCO 'merge instance polygons' option (#938)

* Add polygon merging option to coco converter
* Add test, refactor coco, add support for cli args
* Drop colormap application in datumaro format
* Add cli support in voc converter
* Add cli support in yolo converter
* Add converter cli options in project cli
* Add image data type conversion in image saving

datumaro/datumaro/cli/project/__init__.py

@@ -54,14 +54,14 @@ def build_import_parser(parser):
         help="Source project format (options: %s)" % (', '.join(importers_list)))
     parser.add_argument('-d', '--dest', default='.', dest='dst_dir',
         help="Directory to save the new project to (default: current dir)")
-    parser.add_argument('extra_args', nargs=argparse.REMAINDER,
-        help="Additional arguments for importer")
     parser.add_argument('-n', '--name', default=None,
         help="Name of the new project (default: same as project dir)")
     parser.add_argument('--overwrite', action='store_true',
         help="Overwrite existing files in the save directory")
     parser.add_argument('--copy', action='store_true',
         help="Make a deep copy instead of saving source links")
+    # parser.add_argument('extra_args', nargs=argparse.REMAINDER,
+    #     help="Additional arguments for importer (pass '-- -h' for help)")
     return parser
 
 def import_command(args):
@@ -111,8 +111,8 @@ def build_export_parser(parser):
         help="Output format")
     parser.add_argument('-p', '--project', dest='project_dir', default='.',
         help="Directory of the project to operate on (default: current dir)")
-    parser.add_argument('--save-images', action='store_true',
-        help="Save images")
+    parser.add_argument('extra_args', nargs=argparse.REMAINDER, default=None,
+        help="Additional arguments for converter (pass '-- -h' for help)")
     return parser
 
 def export_command(args):
@@ -125,7 +125,7 @@ def export_command(args):
         save_dir=dst_dir,
         output_format=args.output_format,
         filter_expr=args.filter,
-        save_images=args.save_images)
+        cmdline_args=args.extra_args)
     log.info("Project exported to '%s' as '%s'" % \
         (dst_dir, args.output_format))
 

datumaro/datumaro/components/converter.py

@@ -4,5 +4,16 @@
 # SPDX-License-Identifier: MIT
 
 class Converter:
+    def __init__(self, cmdline_args=None):
+        pass
+
     def __call__(self, extractor, save_dir):
         raise NotImplementedError()
+
+    def _parse_cmdline(self, cmdline):
+        parser = self.build_cmdline_parser()
+
+        if len(cmdline) != 0 and cmdline[0] == '--':
+            cmdline = cmdline[1:]
+        args = parser.parse_args(cmdline)
+        return vars(args)

datumaro/datumaro/components/converters/datumaro.py

@@ -19,7 +19,6 @@ from datumaro.components.extractor import (
 )
 from datumaro.components.formats.datumaro import DatumaroPath
 from datumaro.util.image import save_image
-from datumaro.util.mask_tools import apply_colormap
 
 
 def _cast(value, type_conv, default=None):
@@ -118,13 +117,6 @@ class _SubsetWriter:
         if mask is None:
             return mask_id
 
-        if self._converter._apply_colormap:
-            categories = self._converter._extractor.categories()
-            categories = categories[AnnotationType.mask]
-            colormap = categories.colormap
-
-            mask = apply_colormap(mask, colormap)
-
         mask_id = self._next_mask_id
         self._next_mask_id += 1
 
@@ -232,12 +224,10 @@ class _SubsetWriter:
         return converted
 
 class _Converter:
-    def __init__(self, extractor, save_dir,
-            save_images=False, apply_colormap=False):
+    def __init__(self, extractor, save_dir, save_images=False,):
         self._extractor = extractor
         self._save_dir = save_dir
         self._save_images = save_images
-        self._apply_colormap = apply_colormap
 
     def convert(self):
         os.makedirs(self._save_dir, exist_ok=True)
@@ -282,13 +272,27 @@ class _Converter:
         save_image(image_path, image)
 
 class DatumaroConverter(Converter):
-    def __init__(self, save_images=False, apply_colormap=False):
+    def __init__(self, save_images=False, cmdline_args=None):
         super().__init__()
-        self._save_images = save_images
-        self._apply_colormap = apply_colormap
+
+        self._options = {
+            'save_images': save_images,
+        }
+
+        if cmdline_args is not None:
+            self._options.update(self._parse_cmdline(cmdline_args))
+
+    @classmethod
+    def build_cmdline_parser(cls, parser=None):
+        import argparse
+        if not parser:
+            parser = argparse.ArgumentParser()
+
+        parser.add_argument('--save-images', action='store_true',
+            help="Save images (default: %(default)s)")
+
+        return parser
 
     def __call__(self, extractor, save_dir):
-        converter = _Converter(extractor, save_dir,
-            apply_colormap=self._apply_colormap,
-            save_images=self._save_images)
+        converter = _Converter(extractor, save_dir, **self._options)
         converter.convert()

datumaro/datumaro/components/converters/ms_coco.py

@@ -14,7 +14,7 @@ from datumaro.components.converter import Converter
 from datumaro.components.extractor import (
     DEFAULT_SUBSET_NAME, AnnotationType, PointsObject, BboxObject
 )
-from datumaro.components.formats.ms_coco import CocoAnnotationType, CocoPath
+from datumaro.components.formats.ms_coco import CocoTask, CocoPath
 from datumaro.util import find
 from datumaro.util.image import save_image
 import datumaro.util.mask_tools as mask_tools
@@ -29,8 +29,9 @@ def _cast(value, type_conv, default=None):
         return default
 
 class _TaskConverter:
-    def __init__(self):
+    def __init__(self, context):
         self._min_ann_id = 1
+        self._context = context
 
         data = {
             'licenses': [],
@@ -191,6 +192,13 @@ class _InstancesConverter(_TaskConverter):
                     rle = mask_utils.merge(rles)
                     area = mask_utils.area(rle)
 
+                    if self._context._merge_polygons:
+                        binary_mask = mask_utils.decode(rle).astype(np.bool)
+                        binary_mask = np.asfortranarray(binary_mask, dtype=np.uint8)
+                        segmentation = mask_tools.convert_mask_to_rle(binary_mask)
+                        is_crowd = True
+                        bbox = [int(i) for i in mask_utils.toBbox(rle)]
+
             if ann.group is not None:
                 # Mark the group as visited to prevent repeats
                 for a in annotations[:]:
@@ -201,6 +209,18 @@ class _InstancesConverter(_TaskConverter):
                 is_crowd = False
                 segmentation = [ann.get_polygon()]
                 area = ann.area()
+
+                if self._context._merge_polygons:
+                    h, w, _ = item.image.shape
+                    rles = mask_utils.frPyObjects(segmentation, h, w)
+                    rle = mask_utils.merge(rles)
+                    area = mask_utils.area(rle)
+                    binary_mask = mask_utils.decode(rle).astype(np.bool)
+                    binary_mask = np.asfortranarray(binary_mask, dtype=np.uint8)
+                    segmentation = mask_tools.convert_mask_to_rle(binary_mask)
+                    is_crowd = True
+                    bbox = [int(i) for i in mask_utils.toBbox(rle)]
+
             if bbox is None:
                 bbox = ann.get_bbox()
 
@@ -340,22 +360,30 @@ class _LabelsConverter(_TaskConverter):
 
 class _Converter:
     _TASK_CONVERTER = {
-        CocoAnnotationType.image_info: _ImageInfoConverter,
-        CocoAnnotationType.instances: _InstancesConverter,
-        CocoAnnotationType.person_keypoints: _KeypointsConverter,
-        CocoAnnotationType.captions: _CaptionsConverter,
-        CocoAnnotationType.labels: _LabelsConverter,
+        CocoTask.image_info: _ImageInfoConverter,
+        CocoTask.instances: _InstancesConverter,
+        CocoTask.person_keypoints: _KeypointsConverter,
+        CocoTask.captions: _CaptionsConverter,
+        CocoTask.labels: _LabelsConverter,
     }
 
-    def __init__(self, extractor, save_dir, save_images=False, task=None):
-        if not task:
-            task = list(self._TASK_CONVERTER.keys())
-        elif task in CocoAnnotationType:
-            task = [task]
-        self._task = task
+    def __init__(self, extractor, save_dir,
+            tasks=None, save_images=False, merge_polygons=False):
+        assert tasks is None or isinstance(tasks, (CocoTask, list))
+        if tasks is None:
+            tasks = list(self._TASK_CONVERTER)
+        elif isinstance(tasks, CocoTask):
+            tasks = [tasks]
+        else:
+            for t in tasks:
+                assert t in CocoTask
+        self._tasks = tasks
+
         self._extractor = extractor
         self._save_dir = save_dir
+
         self._save_images = save_images
+        self._merge_polygons = merge_polygons
 
     def make_dirs(self):
         self._images_dir = osp.join(self._save_dir, CocoPath.IMAGES_DIR)
@@ -365,11 +393,13 @@ class _Converter:
         os.makedirs(self._ann_dir, exist_ok=True)
 
     def make_task_converter(self, task):
-        return self._TASK_CONVERTER[task]()
+        if task not in self._TASK_CONVERTER:
+            raise NotImplementedError()
+        return self._TASK_CONVERTER[task](self)
 
     def make_task_converters(self):
         return {
-            task: self.make_task_converter(task) for task in self._task
+            task: self.make_task_converter(task) for task in self._tasks
         }
 
     def save_image(self, item, filename):
@@ -411,32 +441,56 @@ class _Converter:
                         '%s_%s.json' % (task.name, subset_name)))
 
 class CocoConverter(Converter):
-    def __init__(self, task=None, save_images=False):
+    def __init__(self,
+            tasks=None, save_images=False, merge_polygons=False,
+            cmdline_args=None):
         super().__init__()
-        self._task = task
-        self._save_images = save_images
+
+        self._options = {
+            'tasks': tasks,
+            'save_images': save_images,
+            'merge_polygons': merge_polygons,
+        }
+
+        if cmdline_args is not None:
+            self._options.update(self._parse_cmdline(cmdline_args))
+
+    @staticmethod
+    def _split_tasks_string(s):
+        return [CocoTask[i.strip()] for i in s.split(',')]
+
+    @classmethod
+    def build_cmdline_parser(cls, parser=None):
+        import argparse
+        if not parser:
+            parser = argparse.ArgumentParser()
+
+        parser.add_argument('--save-images', action='store_true',
+            help="Save images (default: %(default)s)")
+        parser.add_argument('--merge-polygons', action='store_true',
+            help="Merge instance polygons into a mask (default: %(default)s)")
+        parser.add_argument('--tasks', type=cls._split_tasks_string,
+            default=None,
+            help="COCO task filter, comma-separated list of {%s} "
+                "(default: all)" % ', '.join([t.name for t in CocoTask]))
+
+        return parser
 
     def __call__(self, extractor, save_dir):
-        converter = _Converter(extractor, save_dir,
-            save_images=self._save_images, task=self._task)
+        converter = _Converter(extractor, save_dir, **self._options)
         converter.convert()
 
-def CocoInstancesConverter(save_images=False):
-    return CocoConverter(CocoAnnotationType.instances,
-        save_images=save_images)
+def CocoInstancesConverter(**kwargs):
+    return CocoConverter(CocoTask.instances, **kwargs)
 
-def CocoImageInfoConverter(save_images=False):
-    return CocoConverter(CocoAnnotationType.image_info,
-        save_images=save_images)
+def CocoImageInfoConverter(**kwargs):
+    return CocoConverter(CocoTask.image_info, **kwargs)
 
-def CocoPersonKeypointsConverter(save_images=False):
-    return CocoConverter(CocoAnnotationType.person_keypoints,
-        save_images=save_images)
+def CocoPersonKeypointsConverter(**kwargs):
+    return CocoConverter(CocoTask.person_keypoints, **kwargs)
 
-def CocoCaptionsConverter(save_images=False):
-    return CocoConverter(CocoAnnotationType.captions,
-        save_images=save_images)
+def CocoCaptionsConverter(**kwargs):
+    return CocoConverter(CocoTask.captions, **kwargs)
 
-def CocoLabelsConverter(save_images=False):
-    return CocoConverter(CocoAnnotationType.labels,
-        save_images=save_images)
+def CocoLabelsConverter(**kwargs):
+    return CocoConverter(CocoTask.labels, **kwargs)

datumaro/datumaro/components/converters/voc.py

@@ -31,11 +31,18 @@ class _Converter:
     _BODY_PARTS = set([entry.name for entry in VocBodyPart])
     _ACTIONS = set([entry.name for entry in VocAction])
 
-    def __init__(self, task, extractor, save_dir,
-            apply_colormap=True, save_images=False):
+    def __init__(self, extractor, save_dir,
+            tasks=None, apply_colormap=True, save_images=False):
+        assert tasks is None or isinstance(tasks, (VocTask, list))
+        if tasks is None:
+            tasks = list(VocTask)
+        elif isinstance(tasks, VocTask):
+            tasks = [tasks]
+        else:
+            for t in tasks:
+                assert t in VocTask
+        self._tasks = tasks
 
-        assert not task or task in VocTask
-        self._task = task
         self._extractor = extractor
         self._save_dir = save_dir
         self._apply_colormap = apply_colormap
@@ -205,10 +212,10 @@ class _Converter:
 
                                 objects_with_actions[new_obj_id][action] = presented
 
-                    if self._task in [None,
+                    if set(self._tasks) & set([None,
                             VocTask.detection,
                             VocTask.person_layout,
-                            VocTask.action_classification]:
+                            VocTask.action_classification]):
                         with open(osp.join(self._ann_dir, item_id + '.xml'), 'w') as f:
                             f.write(ET.tostring(root_elem,
                                 encoding='unicode', pretty_print=True))
@@ -245,19 +252,19 @@ class _Converter:
                     action_list[item_id] = None
                     segm_list[item_id] = None
 
-                if self._task in [None,
+                if set(self._tasks) & set([None,
                         VocTask.classification,
                         VocTask.detection,
                         VocTask.action_classification,
-                        VocTask.person_layout]:
+                        VocTask.person_layout]):
                     self.save_clsdet_lists(subset_name, clsdet_list)
-                    if self._task in [None, VocTask.classification]:
+                    if set(self._tasks) & set([None, VocTask.classification]):
                         self.save_class_lists(subset_name, class_lists)
-                if self._task in [None, VocTask.action_classification]:
+                if set(self._tasks) & set([None, VocTask.action_classification]):
                     self.save_action_lists(subset_name, action_list)
-                if self._task in [None, VocTask.person_layout]:
+                if set(self._tasks) & set([None, VocTask.person_layout]):
                     self.save_layout_lists(subset_name, layout_list)
-                if self._task in [None, VocTask.segmentation]:
+                if set(self._tasks) & set([None, VocTask.segmentation]):
                     self.save_segm_lists(subset_name, segm_list)
 
     def save_action_lists(self, subset_name, action_list):
@@ -337,34 +344,57 @@ class _Converter:
         save_image(path, data)
 
 class VocConverter(Converter):
-    def __init__(self, task=None, save_images=False, apply_colormap=False):
+    def __init__(self,
+            tasks=None, save_images=False, apply_colormap=False,
+            cmdline_args=None):
         super().__init__()
-        self._task = task
-        self._save_images = save_images
-        self._apply_colormap = apply_colormap
+
+        self._options = {
+            'tasks': tasks,
+            'save_images': save_images,
+            'apply_colormap': apply_colormap,
+        }
+
+        if cmdline_args is not None:
+            self._options.update(self._parse_cmdline(cmdline_args))
+
+    @staticmethod
+    def _split_tasks_string(s):
+        return [VocTask[i.strip()] for i in s.split(',')]
+
+    @classmethod
+    def build_cmdline_parser(cls, parser=None):
+        import argparse
+        if not parser:
+            parser = argparse.ArgumentParser()
+
+        parser.add_argument('--save-images', action='store_true',
+            help="Save images (default: %(default)s)")
+        parser.add_argument('--apply-colormap', type=bool, default=True,
+            help="Use colormap for class and instance masks "
+                "(default: %(default)s)")
+        parser.add_argument('--tasks', type=cls._split_tasks_string,
+            default=None,
+            help="VOC task filter, comma-separated list of {%s} "
+                "(default: all)" % ', '.join([t.name for t in VocTask]))
+
+        return parser
 
     def __call__(self, extractor, save_dir):
-        converter = _Converter(self._task, extractor, save_dir,
-            apply_colormap=self._apply_colormap,
-            save_images=self._save_images)
+        converter = _Converter(extractor, save_dir, **self._options)
         converter.convert()
 
-def VocClassificationConverter(save_images=False):
-    return VocConverter(VocTask.classification,
-        save_images=save_images)
+def VocClassificationConverter(**kwargs):
+    return VocConverter(VocTask.classification, **kwargs)
 
-def VocDetectionConverter(save_images=False):
-    return VocConverter(VocTask.detection,
-        save_images=save_images)
+def VocDetectionConverter(**kwargs):
+    return VocConverter(VocTask.detection, **kwargs)
 
-def VocLayoutConverter(save_images=False):
-    return VocConverter(VocTask.person_layout,
-        save_images=save_images)
+def VocLayoutConverter(**kwargs):
+    return VocConverter(VocTask.person_layout, **kwargs)
 
-def VocActionConverter(save_images=False):
-    return VocConverter(VocTask.action_classification,
-        save_images=save_images)
+def VocActionConverter(**kwargs):
+    return VocConverter(VocTask.action_classification, **kwargs)
 
-def VocSegmentationConverter(save_images=False, apply_colormap=True):
-    return VocConverter(VocTask.segmentation,
-        save_images=save_images, apply_colormap=apply_colormap)
+def VocSegmentationConverter(**kwargs):
+    return VocConverter(VocTask.segmentation, **kwargs)

datumaro/datumaro/components/converters/yolo.py

@@ -27,11 +27,26 @@ def _make_yolo_bbox(img_size, box):
 class YoloConverter(Converter):
     # https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects
 
-    def __init__(self, task=None, save_images=False, apply_colormap=False):
+    def __init__(self, save_images=False, cmdline_args=None):
         super().__init__()
-        self._task = task
         self._save_images = save_images
-        self._apply_colormap = apply_colormap
+
+        if cmdline_args is not None:
+            options = self._parse_cmdline(cmdline_args)
+            for k, v in options.items():
+                if hasattr(self, '_' + str(k)):
+                    setattr(self, '_' + str(k), v)
+
+    @classmethod
+    def build_cmdline_parser(cls, parser=None):
+        import argparse
+        if not parser:
+            parser = argparse.ArgumentParser()
+
+        parser.add_argument('--save-images', action='store_true',
+            help="Save images (default: %(default)s)")
+
+        return parser
 
     def __call__(self, extractor, save_dir):
         os.makedirs(save_dir, exist_ok=True)

datumaro/datumaro/components/extractors/datumaro.py

@@ -5,6 +5,7 @@
 
 from collections import defaultdict
 import json
+import logging as log
 import os.path as osp
 
 from datumaro.components.extractor import (Extractor, DatasetItem,
@@ -140,11 +141,10 @@ class DatumaroExtractor(Extractor):
                 mask = None
 
                 if osp.isfile(mask_path):
-                    mask_cat = self._categories.get(AnnotationType.mask)
-                    if mask_cat is not None:
-                        mask = lazy_mask(mask_path, mask_cat.inverse_colormap)
-                    else:
-                        mask = lazy_image(mask_path)
+                    mask = lazy_mask(mask_path)
+                else:
+                    log.warn("Not found mask image file '%s', skipped." % \
+                        mask_path)
 
                 loaded.append(MaskObject(label=label_id, image=mask,
                     id=ann_id, attributes=attributes, group=group))

datumaro/datumaro/components/extractors/ms_coco.py

@@ -16,7 +16,7 @@ from datumaro.components.extractor import (Extractor, DatasetItem,
     BboxObject, CaptionObject,
     LabelCategories, PointsCategories
 )
-from datumaro.components.formats.ms_coco import CocoAnnotationType, CocoPath
+from datumaro.components.formats.ms_coco import CocoTask, CocoPath
 from datumaro.util.image import lazy_image
 
 
@@ -103,9 +103,9 @@ class CocoExtractor(Extractor):
 
         self._categories = {}
 
-        label_loader = loaders.get(CocoAnnotationType.labels)
-        instances_loader = loaders.get(CocoAnnotationType.instances)
-        person_kp_loader = loaders.get(CocoAnnotationType.person_keypoints)
+        label_loader = loaders.get(CocoTask.labels)
+        instances_loader = loaders.get(CocoTask.instances)
+        person_kp_loader = loaders.get(CocoTask.person_keypoints)
 
         if label_loader is None and instances_loader is not None:
             label_loader = instances_loader
@@ -209,7 +209,7 @@ class CocoExtractor(Extractor):
         if 'score' in ann:
             attributes['score'] = ann['score']
 
-        if ann_type is CocoAnnotationType.instances:
+        if ann_type is CocoTask.instances:
             x, y, w, h = ann['bbox']
             label_id = self._parse_label(ann)
             group = None
@@ -253,13 +253,13 @@ class CocoExtractor(Extractor):
                 BboxObject(x, y, w, h, label=label_id,
                     id=ann_id, attributes=attributes, group=group)
             )
-        elif ann_type is CocoAnnotationType.labels:
+        elif ann_type is CocoTask.labels:
             label_id = self._parse_label(ann)
             parsed_annotations.append(
                 LabelObject(label=label_id,
                     id=ann_id, attributes=attributes)
             )
-        elif ann_type is CocoAnnotationType.person_keypoints:
+        elif ann_type is CocoTask.person_keypoints:
             keypoints = ann['keypoints']
             points = [p for i, p in enumerate(keypoints) if i % 3 != 2]
             visibility = keypoints[2::3]
@@ -276,7 +276,7 @@ class CocoExtractor(Extractor):
                 parsed_annotations.append(
                     BboxObject(*bbox, label=label_id, group=group)
                 )
-        elif ann_type is CocoAnnotationType.captions:
+        elif ann_type is CocoTask.captions:
             caption = ann['caption']
             parsed_annotations.append(
                 CaptionObject(caption,
@@ -289,21 +289,21 @@ class CocoExtractor(Extractor):
 
 class CocoImageInfoExtractor(CocoExtractor):
     def __init__(self, path, **kwargs):
-        super().__init__(path, task=CocoAnnotationType.image_info, **kwargs)
+        super().__init__(path, task=CocoTask.image_info, **kwargs)
 
 class CocoCaptionsExtractor(CocoExtractor):
     def __init__(self, path, **kwargs):
-        super().__init__(path, task=CocoAnnotationType.captions, **kwargs)
+        super().__init__(path, task=CocoTask.captions, **kwargs)
 
 class CocoInstancesExtractor(CocoExtractor):
     def __init__(self, path, **kwargs):
-        super().__init__(path, task=CocoAnnotationType.instances, **kwargs)
+        super().__init__(path, task=CocoTask.instances, **kwargs)
 
 class CocoPersonKeypointsExtractor(CocoExtractor):
     def __init__(self, path, **kwargs):
-        super().__init__(path, task=CocoAnnotationType.person_keypoints,
+        super().__init__(path, task=CocoTask.person_keypoints,
             **kwargs)
 
 class CocoLabelsExtractor(CocoExtractor):
     def __init__(self, path, **kwargs):
-        super().__init__(path, task=CocoAnnotationType.labels, **kwargs)
+        super().__init__(path, task=CocoTask.labels, **kwargs)

datumaro/datumaro/components/formats/ms_coco.py

@@ -6,11 +6,11 @@
 from enum import Enum
 
 
-CocoAnnotationType = Enum('CocoAnnotationType', [
+CocoTask = Enum('CocoTask', [
     'instances',
     'person_keypoints',
     'captions',
-    'labels', # extension, does not exist in original COCO format
+    'labels', # extension, does not exist in the original COCO format
     'image_info',
     'panoptic',
     'stuff',

datumaro/datumaro/components/importers/ms_coco.py

@@ -7,16 +7,16 @@ from collections import defaultdict
 import os
 import os.path as osp
 
-from datumaro.components.formats.ms_coco import CocoAnnotationType, CocoPath
+from datumaro.components.formats.ms_coco import CocoTask, CocoPath
 
 
 class CocoImporter:
     _COCO_EXTRACTORS = {
-        CocoAnnotationType.instances: 'coco_instances',
-        CocoAnnotationType.person_keypoints: 'coco_person_kp',
-        CocoAnnotationType.captions: 'coco_captions',
-        CocoAnnotationType.labels: 'coco_labels',
-        CocoAnnotationType.image_info: 'coco_images',
+        CocoTask.instances: 'coco_instances',
+        CocoTask.person_keypoints: 'coco_person_kp',
+        CocoTask.captions: 'coco_captions',
+        CocoTask.labels: 'coco_labels',
+        CocoTask.image_info: 'coco_images',
     }
 
     def __init__(self, task_filter=None):
@@ -58,12 +58,12 @@ class CocoImporter:
             name_parts = osp.splitext(ann_file)[0].rsplit('_', maxsplit=1)
             ann_type = name_parts[0]
             try:
-                ann_type = CocoAnnotationType[ann_type]
+                ann_type = CocoTask[ann_type]
             except KeyError:
                 raise Exception(
                     'Unknown subset type %s, only known are: %s' % \
                     (ann_type,
-                     ', '.join([e.name for e in CocoAnnotationType])
+                     ', '.join([e.name for e in CocoTask])
                     ))
             subset_name = name_parts[1]
             subsets[subset_name][ann_type] = subset_path

datumaro/datumaro/components/project.py

@@ -542,7 +542,7 @@ class ProjectDataset(Extractor):
         return self
 
     def save(self, save_dir=None, merge=False, recursive=True,
-            save_images=False, apply_colormap=True):
+            save_images=False):
         if save_dir is None:
             assert self.config.project_dir
             save_dir = self.config.project_dir
@@ -562,7 +562,6 @@ class ProjectDataset(Extractor):
 
         converter_kwargs = {
             'save_images': save_images,
-            'apply_colormap': apply_colormap,
         }
 
         if merge:

datumaro/datumaro/util/image.py

@@ -28,7 +28,7 @@ def load_image(path):
 
     if _IMAGE_BACKEND == _IMAGE_BACKENDS.cv2:
         import cv2
-        image = cv2.imread(path)
+        image = cv2.imread(path, cv2.IMREAD_UNCHANGED)
         image = image.astype(np.float32)
     elif _IMAGE_BACKEND == _IMAGE_BACKENDS.PIL:
         from PIL import Image
@@ -39,13 +39,15 @@ def load_image(path):
     else:
         raise NotImplementedError()
 
-    assert len(image.shape) == 3
-    assert image.shape[2] in [1, 3, 4]
+    assert len(image.shape) in [2, 3]
+    if len(image.shape) == 3:
+        assert image.shape[2] in [3, 4]
     return image
 
 def save_image(path, image, params=None):
     if _IMAGE_BACKEND == _IMAGE_BACKENDS.cv2:
         import cv2
+        image = image.astype(np.uint8)
         cv2.imwrite(path, image, params=params)
     elif _IMAGE_BACKEND == _IMAGE_BACKENDS.PIL:
         from PIL import Image
@@ -109,8 +111,9 @@ def decode_image(image_bytes):
     else:
         raise NotImplementedError()
 
-    assert len(image.shape) == 3
-    assert image.shape[2] in [1, 3, 4]
+    assert len(image.shape) in [2, 3]
+    if len(image.shape) == 3:
+        assert image.shape[2] in [3, 4]
     return image
 
 

datumaro/tests/test_coco_format.py

@@ -138,9 +138,8 @@ class CocoImporterTest(TestCase):
             self.assertFalse(ann_2_mask is None)
 
 class CocoConverterTest(TestCase):
-    def _test_save_and_load(self, source_dataset, converter_type, test_dir,
-            importer_params=None):
-        converter = converter_type()
+    def _test_save_and_load(self, source_dataset, converter, test_dir,
+            importer_params=None, target_dataset=None):
         converter(source_dataset, test_dir.path)
 
         if not importer_params:
@@ -149,6 +148,8 @@ class CocoConverterTest(TestCase):
             **importer_params)
         parsed_dataset = project.make_dataset()
 
+        if target_dataset is not None:
+            source_dataset = target_dataset
         source_subsets = [s if s else DEFAULT_SUBSET_NAME
             for s in source_dataset.subsets()]
         self.assertListEqual(
@@ -195,7 +196,7 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoCaptionsConverter, test_dir)
+                CocoCaptionsConverter(), test_dir)
 
     def test_can_save_and_load_instances(self):
         class TestExtractor(Extractor):
@@ -249,7 +250,7 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoInstancesConverter, test_dir)
+                CocoInstancesConverter(), test_dir)
 
     def test_can_save_and_load_instances_with_mask_conversion(self):
         class TestExtractor(Extractor):
@@ -291,9 +292,64 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoInstancesConverter, test_dir,
+                CocoInstancesConverter(), test_dir,
                 {'merge_instance_polygons': True})
 
+    def test_can_merge_instance_polygons_to_mask_in_coverter(self):
+        label_categories = LabelCategories()
+        for i in range(10):
+            label_categories.add(str(i))
+
+        class SrcTestExtractor(Extractor):
+            def __iter__(self):
+                items = [
+                    DatasetItem(id=0, image=np.zeros((5, 10, 3)),
+                        annotations=[
+                            PolygonObject([0, 0, 4, 0, 4, 4],
+                                label=3, id=4, group=4,
+                                attributes={ 'is_crowd': False }),
+                            PolygonObject([5, 0, 9, 0, 5, 5],
+                                label=3, id=4, group=4,
+                                attributes={ 'is_crowd': False }),
+                        ]
+                    ),
+                ]
+                return iter(items)
+
+            def categories(self):
+                return { AnnotationType.label: label_categories }
+
+        class DstTestExtractor(Extractor):
+            def __iter__(self):
+                items = [
+                    DatasetItem(id=0, image=np.zeros((5, 10, 3)),
+                        annotations=[
+                            BboxObject(1, 0, 8, 4, label=3, id=4, group=4,
+                                attributes={ 'is_crowd': True }),
+                            MaskObject(np.array([
+                                    [0, 1, 1, 1, 0, 1, 1, 1, 1, 0],
+                                    [0, 0, 1, 1, 0, 1, 1, 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]],
+                                    # only internal fragment (without the border),
+                                    # but not everywhere...
+                                    dtype=np.bool),
+                                attributes={ 'is_crowd': True },
+                                label=3, id=4, group=4),
+                        ]
+                    ),
+                ]
+                return iter(items)
+
+            def categories(self):
+                return { AnnotationType.label: label_categories }
+
+        with TestDir() as test_dir:
+            self._test_save_and_load(SrcTestExtractor(),
+                CocoInstancesConverter(merge_polygons=True), test_dir,
+                target_dataset=DstTestExtractor())
+
     def test_can_save_and_load_images(self):
         class TestExtractor(Extractor):
             def __iter__(self):
@@ -314,7 +370,7 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoImageInfoConverter, test_dir)
+                CocoImageInfoConverter(), test_dir)
 
     def test_can_save_and_load_labels(self):
         class TestExtractor(Extractor):
@@ -350,7 +406,7 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoLabelsConverter, test_dir)
+                CocoLabelsConverter(), test_dir)
 
     def test_can_save_and_load_keypoints(self):
         class TestExtractor(Extractor):
@@ -397,7 +453,7 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoPersonKeypointsConverter, test_dir)
+                CocoPersonKeypointsConverter(), test_dir)
 
     def test_can_save_dataset_with_no_subsets(self):
         class TestExtractor(Extractor):
@@ -429,4 +485,4 @@ class CocoConverterTest(TestCase):
 
         with TestDir() as test_dir:
             self._test_save_and_load(TestExtractor(),
-                CocoConverter, test_dir)
+                CocoConverter(), test_dir)

datumaro/tests/test_datumaro_format.py

@@ -18,7 +18,7 @@ class DatumaroConverterTest(TestCase):
     class TestExtractor(Extractor):
         def __iter__(self):
             items = [
-                DatasetItem(id=100, subset='train',
+                DatasetItem(id=100, subset='train', image=np.ones((10, 6, 3)),
                     annotations=[
                         CaptionObject('hello', id=1),
                         CaptionObject('world', id=2, group=5),
@@ -75,8 +75,7 @@ class DatumaroConverterTest(TestCase):
         with TestDir() as test_dir:
             source_dataset = self.TestExtractor()
 
-            converter = DatumaroConverter(
-                save_images=True, apply_colormap=True)
+            converter = DatumaroConverter(save_images=True)
             converter(source_dataset, test_dir.path)
 
             project = Project.import_from(test_dir.path, 'datumaro')

datumaro/tests/test_image.py

@@ -17,9 +17,12 @@ class ImageTest(TestCase):
 
     def test_save_and_load_backends(self):
         backends = image_module._IMAGE_BACKENDS
-        for save_backend, load_backend in product(backends, backends):
+        for save_backend, load_backend, c in product(backends, backends, [1, 3]):
             with TestDir() as test_dir:
-                src_image = np.random.randint(0, 255 + 1, (2, 4, 3))
+                if c == 1:
+                    src_image = np.random.randint(0, 255 + 1, (2, 4))
+                else:
+                    src_image = np.random.randint(0, 255 + 1, (2, 4, c))
                 path = osp.join(test_dir.path, 'img.png') # lossless
 
                 image_module._IMAGE_BACKEND = save_backend
@@ -33,8 +36,11 @@ class ImageTest(TestCase):
 
     def test_encode_and_decode_backends(self):
         backends = image_module._IMAGE_BACKENDS
-        for save_backend, load_backend in product(backends, backends):
-            src_image = np.random.randint(0, 255 + 1, (2, 4, 3))
+        for save_backend, load_backend, c in product(backends, backends, [1, 3]):
+            if c == 1:
+                src_image = np.random.randint(0, 255 + 1, (2, 4))
+            else:
+                src_image = np.random.randint(0, 255 + 1, (2, 4, c))
 
             image_module._IMAGE_BACKEND = save_backend
             buffer = image_module.encode_image(src_image, '.png') # lossless