Add MMDetection COCO format importer

CHANGELOG.md

@@ -6,6 +6,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## \[Unreleased\]
+### New features
+- Support MMDetection COCO format
+  (<https://github.com/openvinotoolkit/datumaro/pull/1213>)
+
 ### Enhancements
 - Optimize Python import to make CLI entrypoint faster
   (<https://github.com/openvinotoolkit/datumaro/pull/1182>)

docs/source/docs/data-formats/formats/index.rst

@@ -32,6 +32,7 @@ Supported Data Formats
    mapillary_vistas
    market1501
    mars
+   mmdet
    mnist
    mot
    mots
@@ -141,6 +142,10 @@ Supported Data Formats
    * `Format specification <https://zheng-lab.cecs.anu.edu.au/Project/project_mars.html>`_
    * `Dataset example <https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/mars_dataset>`_
    * `Format documentation <mars.md>`_
+* MMDet-COCO (``detection``, ``segmentation``)
+   * `Format specification <https://mmdetection.readthedocs.io/en/latest/user_guides/dataset_prepare.html>`_
+   * `Dataset example <https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/coco_dataset/mmdet_coco>`_
+   * `Format documentation <mmdet.md>`_
 * MNIST (``classification``)
    * `Format specification <http://yann.lecun.com/exdb/mnist/>`_
    * `Dataset example <https://github.com/openvinotoolkit/datumaro/tree/develop/tests/assets/mnist_dataset>`_

docs/source/docs/data-formats/formats/mmdet.md

@@ -0,0 +1,41 @@
+# MMDetection COCO
+
+## Format specification
+
+[MMDetection](https://mmdetection.readthedocs.io/en/latest/) is a training framework for object detection and instance segmentation tasks, providing a modular and flexible architecture that supports various state-of-the-art models, datasets, and training techniques. MMDetection has gained popularity in the research community for its comprehensive features and ease of use in developing and benchmarking object detection algorithms.
+MMDetection specifies their COCO format [here](https://mmdetection.readthedocs.io/en/latest/user_guides/dataset_prepare.html).
+
+Most of available tasks or formats are similar to the [original COCO format](./formats/coco), while only the image directories are separated with respect to subsets.
+In this document, we just describe the directory structure of MMDetection COCO format as per [here](https://mmdetection.readthedocs.io/en/latest/user_guides/dataset_prepare.html).
+MMDetection COCO dataset directory should have the following structure:
+
+<!--lint disable fenced-code-flag-->
+```
+└─ Dataset/
+    ├── <subset_name>/
+    │   ├── <image_name1.ext>
+    │   ├── <image_name2.ext>
+    │   └── ...
+    ├── <subset_name>/
+    │   ├── <image_name1.ext>
+    │   ├── <image_name2.ext>
+    │   └── ...
+    └── annotations/
+        ├── instances_<subset_name>.json
+        └── ...
+```
+
+### Import using CLI
+
+``` bash
+datum project create
+datum project import --format mmdet_coco <path/to/dataset>
+```
+
+### Import using Python API
+
+```python
+import datumaro as dm
+
+dataset = dm.Dataset.import_from('<path/to/dataset>', 'mmdet_coco')
+```

src/datumaro/plugins/data_formats/coco/base.py

@@ -88,6 +88,25 @@
         return osp.join(rootpath, subset)
 
 
+class MmdetDirPathExtracter(DirPathExtracter):
+    @staticmethod
+    def find_rootpath(path: str) -> str:
+        """Find root path from annotation json file path."""
+        path = osp.abspath(path)
+        if osp.dirname(path).endswith(CocoPath.ANNOTATIONS_DIR):
+            return path.rsplit(CocoPath.ANNOTATIONS_DIR, maxsplit=1)[0]
+        raise DatasetImportError(
+            f"Annotation path ({path}) should be under the directory which is named {CocoPath.ANNOTATIONS_DIR}. "
+            "If not, Datumaro fails to find the root path for this dataset. "
+            "Please follow this instruction, https://github.com/cocodataset/cocoapi/blob/master/README.txt"
+        )
+
+    @staticmethod
+    def find_images_dir(rootpath: str, subset: str) -> str:
+        """Find images directory from the root path."""
+        return osp.join(rootpath, subset)
+
+
 class _CocoBase(SubsetBase):
     """
     Parses COCO annotations written in the following format:
@@ -121,6 +140,9 @@
         elif coco_importer_type == CocoImporterType.roboflow:
             self._rootpath = RoboflowDirPathExtracter.find_rootpath(path)
             self._images_dir = RoboflowDirPathExtracter.find_images_dir(self._rootpath, subset)
+        elif coco_importer_type == CocoImporterType.mmdet:
+            self._rootpath = MmdetDirPathExtracter.find_rootpath(path)
+            self._images_dir = MmdetDirPathExtracter.find_images_dir(self._rootpath, subset)
         else:
             raise DatasetImportError(f"Not supported type: {coco_importer_type}")
 

src/datumaro/plugins/data_formats/coco/format.py

@@ -18,6 +18,7 @@ class CocoTask(Enum):
 class CocoImporterType(Enum):
     default = auto()
     roboflow = auto()
+    mmdet = auto()
 
 
 class CocoPath:

src/datumaro/plugins/data_formats/mmdet.py

@@ -0,0 +1,79 @@
+# Copyright (C) 2023 Intel Corporation
+#
+# SPDX-License-Identifier: MIT
+
+import os.path as osp
+from glob import glob
+from typing import Optional
+
+from datumaro.components.dataset_base import DEFAULT_SUBSET_NAME
+from datumaro.components.format_detection import FormatDetectionConfidence, FormatDetectionContext
+from datumaro.components.importer import ImportContext
+from datumaro.plugins.data_formats.coco.base import _CocoBase
+from datumaro.plugins.data_formats.coco.format import CocoImporterType, CocoTask
+from datumaro.plugins.data_formats.coco.importer import CocoImporter
+
+
+class MmdetCocoImporter(CocoImporter):
+    @classmethod
+    def detect(
+        cls,
+        context: FormatDetectionContext,
+    ) -> FormatDetectionConfidence:
+        ann_paths = context.require_files("annotations/instances_*.json")
+
+        for ann_path in ann_paths:
+            subset_name = cls._get_subset_name(ann_path)
+
+            with context.require_any():
+                with context.alternative():
+                    image_files = osp.join(subset_name, "*.jpg")
+                    context.require_file(f"{image_files}")
+
+                    return FormatDetectionConfidence.MEDIUM
+
+    def __call__(self, path, stream: bool = False, **extra_params):
+        subset_paths = glob(osp.join(path, "**", "instances_*.json"), recursive=True)
+
+        sources = []
+        for subset_path in subset_paths:
+            options = dict(extra_params)
+            options["subset"] = self._get_subset_name(subset_path)
+
+            if stream:
+                options["stream"] = True
+
+            sources.append({"url": subset_path, "format": "mmdet_coco", "options": options})
+
+        return sources
+
+    @classmethod
+    def _get_subset_name(cls, subset_path: str):
+        parts = osp.splitext(osp.basename(subset_path))[0].split("instances_", maxsplit=1)
+        subset_name = parts[1] if len(parts) == 2 else DEFAULT_SUBSET_NAME
+
+        return subset_name
+
+
+class MmdetCocoBase(_CocoBase):
+    """
+    Parses Roboflow COCO annotations written in the following format:
+    https://cocodataset.org/#format-data
+    """
+
+    def __init__(
+        self,
+        path,
+        *,
+        subset: Optional[str] = None,
+        stream: bool = False,
+        ctx: Optional[ImportContext] = None,
+    ):
+        super().__init__(
+            path,
+            task=CocoTask.instances,
+            coco_importer_type=CocoImporterType.mmdet,
+            subset=subset,
+            stream=stream,
+            ctx=ctx,
+        )

src/datumaro/plugins/specs.json

@@ -710,6 +710,18 @@
     "plugin_type": "Importer",
     "extra_deps": []
   },
+  {
+    "import_path": "datumaro.plugins.data_formats.mmdet.MmdetCocoBase",
+    "plugin_name": "mmdet_coco",
+    "plugin_type": "DatasetBase",
+    "extra_deps": []
+  },
+  {
+    "import_path": "datumaro.plugins.data_formats.mmdet.MmdetCocoImporter",
+    "plugin_name": "mmdet_coco",
+    "plugin_type": "Importer",
+    "extra_deps": []
+  },
   {
     "import_path": "datumaro.plugins.data_formats.mnist.MnistBase",
     "plugin_name": "mnist",

tests/assets/coco_dataset/mmdet_coco/annotations/instances_train.json

@@ -0,0 +1,64 @@
+{
+    "licenses":[
+      {
+        "name":"",
+        "id":0,
+        "url":""
+      }
+    ],
+    "info":{
+      "contributor":"",
+      "date_created":"",
+      "description":"",
+      "url":"",
+      "version":"",
+      "year":""
+    },
+    "categories":[
+      {
+        "id":1,
+        "name":"a",
+        "supercategory":""
+      },
+      {
+        "id":2,
+        "name":"b",
+        "supercategory":""
+      },
+      {
+        "id":4,
+        "name":"c",
+        "supercategory":""
+      }
+    ],
+    "images":[
+      {
+        "id":5,
+        "width":10,
+        "height":5,
+        "file_name":"a.jpg",
+        "license":0,
+        "flickr_url":"",
+        "coco_url":"",
+        "date_captured":0
+      }
+    ],
+    "annotations":[
+      {
+        "id":1,
+        "image_id":5,
+        "category_id":2,
+        "segmentation":[
+
+        ],
+        "area":3.0,
+        "bbox":[
+          2.0,
+          2.0,
+          3.0,
+          1.0
+        ],
+        "iscrowd":0
+      }
+    ]
+  }

tests/assets/coco_dataset/mmdet_coco/annotations/instances_val.json

@@ -0,0 +1,101 @@
+{
+    "licenses":[
+      {
+        "name":"",
+        "id":0,
+        "url":""
+      }
+    ],
+    "info":{
+      "contributor":"",
+      "date_created":"",
+      "description":"",
+      "url":"",
+      "version":"",
+      "year":""
+    },
+    "categories":[
+      {
+        "id":1,
+        "name":"a",
+        "supercategory":""
+      },
+      {
+        "id":2,
+        "name":"b",
+        "supercategory":""
+      },
+      {
+        "id":4,
+        "name":"c",
+        "supercategory":""
+      }
+    ],
+    "images":[
+      {
+        "id":40,
+        "width":5,
+        "height":10,
+        "file_name":"b.jpg",
+        "license":0,
+        "flickr_url":"",
+        "coco_url":"",
+        "date_captured":0
+      }
+    ],
+    "annotations":[
+      {
+        "id":1,
+        "image_id":40,
+        "category_id":1,
+        "segmentation":[
+          [
+            0.0,
+            0.0,
+            1.0,
+            0.0,
+            1.0,
+            2.0,
+            0.0,
+            2.0
+          ]
+        ],
+        "area":2.0,
+        "bbox":[
+          0.0,
+          0.0,
+          1.0,
+          2.0
+        ],
+        "iscrowd":0,
+        "attributes":{
+          "x":1,
+          "y":"hello"
+        }
+      },
+      {
+        "id":2,
+        "image_id":40,
+        "category_id":2,
+        "segmentation":{
+          "counts":[
+            0,
+            20,
+            30
+          ],
+          "size":[
+            10,
+            5
+          ]
+        },
+        "area":20.0,
+        "bbox":[
+          0.0,
+          0.0,
+          1.0,
+          9.0
+        ],
+        "iscrowd":1
+      }
+    ]
+  }