Import for SYNTHIA dataset format

CHANGELOG.md

@@ -12,6 +12,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   (<https://github.com/openvinotoolkit/datumaro/pull/537>)
 - Datumaro can now be installed on Windows on Python 3.9
   (<https://github.com/openvinotoolkit/datumaro/pull/547>)
+- Import for SYNTHIA dataset format.
+  (<https://github.com/openvinotoolkit/datumaro/pull/532>)
 
 ### Changed
 - The following formats can now be detected unambiguously:

datumaro/plugins/synthia_format.py

@@ -0,0 +1,180 @@
+# Copyright (C) 2021 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+from collections import OrderedDict
+import os.path as osp
+
+import numpy as np
+
+from datumaro.components.annotation import (
+    AnnotationType, LabelCategories, Mask, MaskCategories,
+)
+from datumaro.components.extractor import DatasetItem, Importer, SourceExtractor
+from datumaro.components.format_detection import FormatDetectionContext
+from datumaro.util.image import find_images, load_image
+from datumaro.util.mask_tools import generate_colormap, lazy_mask
+
+
+class SynthiaPath:
+    IMAGES_DIR = 'RGB'
+    LABELS_SEGM_DIR = 'GT/LABELS'
+    SEMANTIC_SEGM_DIR = 'GT/COLOR'
+    LABELMAP_FILE = 'label_colors.txt'
+
+SYNTHIA_LABEL_MAP = OrderedDict([
+    ('Void', (0, 0, 0)),
+    ('Sky', (128, 128, 128)),
+    ('Building', (128, 0, 0)),
+    ('Road', (128, 64, 128)),
+    ('Sidewalk', (0, 0, 192)),
+    ('Fence', (64, 64, 128)),
+    ('Vegetation', (128, 128, 0)),
+    ('Pole', (192, 192, 128)),
+    ('Car', (64, 0, 128)),
+    ('TrafficSign', (192, 128, 128)),
+    ('Pedestrian', (64, 64, 0)),
+    ('Bicycle', (0, 128, 192)),
+    ('Lanemarking', (0, 172, 0)),
+    ('Reserved_1', (0, 0, 0)),
+    ('Reserved_2', (0, 0, 0)),
+    ('TrafficLight', (0, 128, 128)),
+])
+
+def make_categories(label_map=None):
+    if label_map is None:
+        label_map = SYNTHIA_LABEL_MAP
+
+    categories = {}
+    label_categories = LabelCategories()
+    for label in label_map:
+        label_categories.add(label)
+    categories[AnnotationType.label] = label_categories
+
+    has_colors = any(v is not None for v in label_map.values())
+    if not has_colors: # generate new colors
+        colormap = generate_colormap(len(label_map))
+    else: # only copy defined colors
+        colormap = { label_id: (desc[0], desc[1], desc[2])
+            for label_id, desc in enumerate(label_map.values()) }
+    mask_categories = MaskCategories(colormap)
+    mask_categories.inverse_colormap # pylint: disable=pointless-statement
+    categories[AnnotationType.mask] = mask_categories
+    return categories
+
+def parse_label_map(path):
+    label_map = OrderedDict()
+    with open(path, 'r', encoding='utf-8') as f:
+        for line in f:
+            # skip empty and commented lines
+            line = line.strip()
+            if not line or line[0] == '#':
+                continue
+
+            # color, name
+            label_desc = line.split()
+
+            if 2 < len(label_desc):
+                name = label_desc[3]
+                color = tuple([int(c) for c in label_desc[:3]])
+            else:
+                name = label_desc[0]
+                color = None
+
+            if name in label_map:
+                raise ValueError("Label '%s' is already defined" % name)
+
+            label_map[name] = color
+    return label_map
+
+class SynthiaExtractor(SourceExtractor):
+    def __init__(self, path):
+        if not osp.isdir(path):
+            raise FileNotFoundError("Can't read dataset directory '%s'" % path)
+
+        super().__init__()
+
+        self._categories = self._load_categories(path)
+        self._items = list(self._load_items(path).values())
+
+    def _load_categories(self, path):
+        label_map_path = osp.join(path, SynthiaPath.LABELMAP_FILE)
+        if osp.isfile(label_map_path):
+            label_map = parse_label_map(label_map_path)
+        else:
+            label_map = SYNTHIA_LABEL_MAP
+        return make_categories(label_map)
+
+    def _load_items(self, root_dir):
+        image_dir = osp.join(root_dir, SynthiaPath.IMAGES_DIR)
+        if osp.isdir(image_dir):
+            images = {
+                osp.splitext(osp.relpath(p, image_dir))[0].replace('\\', '/'): p
+                for p in find_images(image_dir, recursive=True)
+            }
+        else:
+            images = {}
+
+        items = {}
+
+        inst_dir = osp.join(root_dir, SynthiaPath.LABELS_SEGM_DIR)
+        if osp.isdir(inst_dir):
+            gt_images = find_images(inst_dir, recursive=True)
+            for gt_img in gt_images:
+                item_id = osp.splitext(osp.relpath(gt_img, inst_dir))[0].replace('\\', '/')
+
+                anno = []
+                labels_mask = load_image(gt_img, dtype=np.uint16)
+                dynamic_objects = np.unique(labels_mask[:,:,1])
+                labels_mask = labels_mask[:,:,2]
+                segm_ids = np.unique(labels_mask)
+                for segm_id in segm_ids:
+                    attr = { 'dynamic_object': False }
+                    if segm_id in dynamic_objects:
+                        attr['dynamic_object'] = True
+                    anno.append(Mask(
+                        image=self._lazy_extract_mask(labels_mask, segm_id),
+                        label=segm_id, attributes=attr))
+
+                items[item_id] = DatasetItem(id=item_id, image=images[item_id],
+                    annotations=anno)
+
+        elif osp.isdir(osp.join(root_dir, SynthiaPath.SEMANTIC_SEGM_DIR)):
+            gt_dir = osp.join(root_dir, SynthiaPath.SEMANTIC_SEGM_DIR)
+            gt_images = find_images(gt_dir, recursive=True)
+            for gt_img in gt_images:
+                item_id = osp.splitext(osp.relpath(gt_img, gt_dir))[0].replace('\\', '/')
+
+                anno = []
+                inverse_cls_colormap = \
+                    self._categories[AnnotationType.mask].inverse_colormap
+                color_mask = lazy_mask(gt_img, inverse_cls_colormap)
+                color_mask = color_mask()
+                classes = np.unique(color_mask)
+                for label_id in classes:
+                    anno.append(Mask(image=self._lazy_extract_mask(color_mask, label_id),
+                        label=label_id))
+
+                items[item_id] = DatasetItem(id=item_id, image=images[item_id],
+                    annotations=anno)
+
+
+        return items
+
+    @staticmethod
+    def _lazy_extract_mask(mask, c):
+        return lambda: mask == c
+
+class SynthiaImporter(Importer):
+    @classmethod
+    def detect(cls, context: FormatDetectionContext) -> None:
+        with context.require_any():
+            for prefix in (
+                SynthiaPath.IMAGES_DIR, SynthiaPath.LABELS_SEGM_DIR, SynthiaPath.SEMANTIC_SEGM_DIR
+            ):
+                with context.alternative():
+                    context.require_file(f'{prefix}/**/*.png')
+
+    @classmethod
+    def find_sources(cls, path):
+        return [{'url': path, 'format': 'synthia'}]

site/content/en/docs/formats/synthia.md

@@ -0,0 +1,129 @@
+---
+title: 'SYNTHIA'
+linkTitle: 'SYNTHIA'
+description: ''
+weight: 1
+---
+
+## Format specification
+
+The original SYNTHIA dataset is available
+[here](https://synthia-dataset.net).
+
+Supported annotation types:
+- `Mask`
+
+Supported annotation attributes:
+- `dynamic_object` (boolean): whether the object moving
+
+## Import SYNTHIA dataset
+
+A Datumaro project with a SYNTHIA source can be created in the following way:
+
+```bash
+datum create
+datum import --format synthia <path/to/dataset>
+```
+
+It is also possible to import the dataset using Python API:
+
+```python
+from datumaro.components.dataset import Dataset
+
+synthia_dataset = Dataset.import_from('<path/to/dataset>', 'synthia')
+```
+
+SYNTHIA dataset directory should have the following structure:
+
+<!--lint disable fenced-code-flag-->
+```
+dataset/
+├── GT/
+│   ├── COLOR/
+│   │   ├── Stereo_Left/
+│   │   │   ├── Omni_B
+│   │   │   │   ├── 000000.png
+│   │   │   │   ├── 000001.png
+│   │   │   │   └── ...
+│   │   │   └── ...
+│   │   └── Stereo_Right
+│   │       ├── Omni_B
+│   │       │   ├── 000000.png
+│   │       │   ├── 000001.png
+│   │       │   └── ...
+│   │       └── ...
+│   └── LABELS
+│       ├── Stereo_Left
+│       │   ├── Omni_B
+│       │   │   ├── 000000.png
+│       │   │   ├── 000001.png
+│       │   │   └── ...
+│       │   └── ...
+│       └── Stereo_Right
+│           ├── Omni_B
+│           │   ├── 000000.png
+│           │   ├── 000001.png
+│           │   └── ...
+│           └── ...
+└── RGB
+    ├── Stereo_Left
+    │   ├── Omni_B
+    │   │   ├── 000000.png
+    │   │   ├── 000001.png
+    │   │   └── ...
+    │   └── ...
+    └── Stereo_Right
+        ├── Omni_B
+        │   ├── 000000.png
+        │   ├── 000001.png
+        │   └── ...
+        └── ...
+```
+
+- `RGB` folder containing standard RGB images used for training.
+- `GT/LABELS` folder containing containing PNG files (one per image).
+  Annotations are given in three channels. The red channel contains
+  the class of that pixel. The green channel contains the class only
+  for those objects that are dynamic (cars, pedestrians, etc.),
+  otherwise it contains `0`.
+- `GT/COLOR` folder containing png files (one per image).
+  Annotations are given using a color representation.
+
+When importing a dataset, only `GT/LABELS` folder will be used.
+If it is missing, `GT/COLOR` folder will be used
+
+The original dataset also contains depth information, but Datumaro
+does not currently support it.
+
+
+## Export to other formats
+
+Datumaro can convert a SYNTHIA dataset into any other format [Datumaro supports](/docs/user-manual/supported_formats/).
+To get the expected result, convert the dataset to a format
+that supports segmentation masks.
+
+There are several ways to convert a SYNTHIA dataset to other dataset
+formats using CLI:
+
+```bash
+datum create
+datum import -f synthia <path/to/dataset>
+datum export -f voc -o <output/dir> -- --save-images
+# or
+datum convert -if synthia -i <path/to/dataset> \
+    -f voc -o <output/dir> -- --save-images
+```
+
+Or, using Python API:
+
+```python
+from datumaro.components.dataset import Dataset
+
+dataset = Dataset.import_from('<path/to/dataset>', 'synthia')
+dataset.export('save_dir', 'voc')
+```
+
+## Examples
+
+Examples of using this format from the code can be found in
+[the format tests](https://github.com/openvinotoolkit/datumaro/blob/develop/tests/test_synthia_format.py)

site/content/en/docs/user-manual/supported_formats.md

@@ -73,6 +73,10 @@ List of supported formats:
   - [Format specification](https://docs.supervise.ly/data-organization/00_ann_format_navi)
   - [Dataset example](https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/sly_pointcloud_dataset)
   - [Format documentation](/docs/formats/sly_pointcloud)
+- SYNTHIA (`segmentation`)
+  - [Format specification](https://synthia-dataset.net/)
+  - [Dataset example](https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/synthia_dataset)
+  - [Format documentation](/docs/formats/synthia)
 - CVAT
   - [Format specification](https://openvinotoolkit.github.io/cvat/docs/manual/advanced/xml_format)
   - [Dataset example](https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/cvat_dataset)

tests/assets/synthia_dataset/dataset_with_custom_labelmap/label_colors.txt

@@ -0,0 +1,5 @@
+0 0 0 background
+0 0 64 sky
+0 128 128 building
+128 0 64 person
+0 192 128 road

tests/requirements.py

@@ -27,6 +27,7 @@ class Requirements:
     DATUM_283 = "Create cli tests for testing convert command for VOC format"
     DATUM_399 = "Implement import for ADE20K dataset"
     DATUM_475 = "Support import for CelebA dataset"
+    DATUM_497 = "Support import for SYNTHIA dataset"
 
     # GitHub issues (bugs)
     # https://github.com/openvinotoolkit/datumaro/issues