Supported Data Formats

• ADE20k (v2017) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• ADE20k (v2020) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• Align CelebA (classification, landmarks) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• BraTS (segmentation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• BraTS Numpy (detection, segmentation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• CamVid (segmentation)

  • Format specification

  • Dataset example

• CelebA (classification, detection, landmarks) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• CIFAR-10/100 (classification)

  • Format specification

  • Dataset example CIFAR-10

  • Dataset example CIFAR-100

  • Format documentation

• Cityscapes (segmentation)

  • Format specification

  • Dataset example

  • Format documentation

• Common Semantic Segmentation (segmentation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• Common Super Resolution

  • Format specification

  • Dataset example

  • Format documentation

• CVAT (for images, for video (import-only))

  • Format specification

  • Dataset example

  • Format documentation

• DOTA (detection_rotated)

  • Format specification

  • Dataset example

  • Format documentation

• ICDAR13/15 (word recognition, text localization, text segmentation)

  • Format specification

  • Dataset example

• ImageNet (classification, detection)

  • Dataset example

  • Dataset example (txt for classification)

  • Detection format is the same as in PASCAL VOC

  • Format documentation

• Kaggle (classification, detection, segmentation) (import-only)

  • Dataset examples

  • Format documentation

• KITTI (segmentation, detection)

  • Format specification

  • Dataset example

  • Format documentation

• KITTI 3D (raw, tracklets, velodyne points)

  • Format specification

  • Dataset example

  • Format documentation

• Kinetics 400/600/700

  • Format specification

  • Dataset example

  • Format documentation

• LabelMe (labels, boxes, masks)

  • Format specification

  • Dataset example

  • Format documentation

• LFW (classification, person re-identification, landmarks)

  • Format specification

  • Dataset example

  • Format documentation

• Mapillary Vistas (segmentation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• Market-1501 (person re-identification)

  • Format specification

  • Dataset example

• MARS (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• MMDet-COCO (detection, segmentation)

  • Format specification

  • Dataset example

  • Format documentation

• MNIST (classification)

  • Format specification

  • Dataset example

  • Format documentation

• MNIST in CSV (classification)

  • Format specification

  • Dataset example

  • Format documentation

• MOT sequences

  • Format specification

  • Dataset example

  • Format documentation

• MOTS (png)

  • Format specification

  • Dataset example

  • Format documentation

• MPII Human Pose (detection, pose estimation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• MPII Human Pose JSON (detection, pose estimation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• MS COCO (image info, instances, person keypoints, captions, labels, panoptic, stuff)

  • Format specification

  • Dataset example

  • labels are our extension - like instances with only category_id

  • Format documentation

• Roboflow (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• NYU Depth Dataset V2 (depth estimation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• OpenImages (classification, detection, segmentation)

  • Format specification

  • Dataset example

  • Format documentation

• PASCAL VOC (classification, detection, segmentation, action classification, person layout)

  • Format specification

  • Dataset example

  • Format documentation

• Segment Anything (a.k.a SA-1B) (detection, segmentation)

  • Format specification

  • Dataset example

  • Format documentation

• Supervisely (pointcloud)

  • Format specification

  • Dataset example

  • Format documentation

• SYNTHIA (segmentation) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• Tabular (classification, regression) (import/export only)

  • Dataset example

  • Format documentation

• TF Detection API (bboxes, masks)

  • Format specifications: [bboxes], [masks]

  • Dataset example

• VGGFace2 (landmarks, bboxes)

  • Format specification

  • Dataset example

  • Format documentation

• VoTT CSV (detection) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• VoTT JSON (detection) (import-only)

  • Format specification

  • Dataset example

  • Format documentation

• WIDERFace (bboxes)

  • Format specification

  • Dataset example

  • Format documentation

• YOLO (bboxes)

  • Format specification

  • Dataset example

  • Format documentation

• YOLO-Ultralytics (bboxes)

  • Format specification

  • Dataset example

  • Format documentation

Supported Annotation Types

• Labels

• Bounding Boxes

• Polygons

• Polylines

• (Segmentation) Masks

• (Key-) Points

• Captions

• 3D cuboids

• Super Resolution Annotation

• Depth Annotation

• Ellipses

• Hash Keys

Datumaro does not separate datasets by tasks like classification, detection, segmentation, etc. Instead, datasets can have any annotations. When a dataset is exported in a specific format, only relevant annotations are exported.

Dataset Meta Info File

It is possible to use classes that are not original to the format. To do this, use dataset_meta.json.

{
  "label_map": {"0": "background", "1": "car", "2": "person"},
  "segmentation_colors": [[0, 0, 0], [255, 0, 0], [0, 0, 255]],
  "background_label": "0"
}

• label_map is a dictionary where the class ID is the key and the class name is the value.

• segmentation_colors is a list of channel-wise values for each class. This is only necessary for the segmentation task.

• background_label is a background label ID in the dataset.