For the complete documentation index, see llms.txt. This page is also available as Markdown.

Object detection format

Description of the Mindkosh annotation format for object detection in images.

Image annotations in Mindkosh format contains annotations grouped by the image they are in, and the annotation type, which can be bounding boxes, polygons, polylines and points (keypoints). {

"annotations": {
    "version": "1.1",
    "meta": {
      "task": {
        "labels": [
          {
            "id": 11695,
            "name": "goal_post",
            "attributes": [
              {
                "id": 1757,
                "name": "occlusion",
                "mutable": true,
                "input_type": "checkbox",
                "default_value": "",
                "values": "true\nfalse"
              }
            ]
          }
        ]
      }
    },
    "images": [
      {
        "id": 0,
        "name": "psg_marseille_first_half_1228.png",
        "width": 1280,
        "height": 720,
        "boxes": [
          {
            "id": 3015293,
            "label": 11695,
            "source": "manual",
            "point_order": -1,
            "related_ids": [],
            "created_mode": "annotation",
            "created_by": "sdevgupta",
            "updated_by": "None",
            "xtl": "599.73",
            "ytl": "163.63",
            "xbr": "658.33",
            "ybr": "188.35",
            "rotation": "21.3",
            "z_order": 0,
            "attributes": [
              { "id": 1757, "name": "occlusion", "value": "true" }
            ],
          }
        ]
      }
    ]
  }
}

Property Description

id

Annotation ID. Unique across Mindkosh.

label

Label ID. Can be referenced from the labels section

source

Can be manual, imported or automatic

related_ids

Array of annotation ID of other objects connected with this annotation

created_mode

Which annotation mode this annotation was created in. Can be annotation, validation, qc or completed

created_by

User who created the annotation

updated_by

User who updated the annotation

attributes

key value pairs of attributes if available

rotation

Orientation of a rotated bounding box. Excluded for axis-parallel bounding boxes.

z_order

Layer number of an annotation. Only applicable to polygons

Annotation description

boxes

Described by the top left and bottom right points of the box. xtl, ytl, xbr, ybr

polygons

Flat array of the points - x1, y1, x2, y2, x3, y3. Last point connects to the first point.

polylines

Same as the polygons. Last point does not connect to the first point.

points

Flat array of the points

cuboids

Described by the top-left and bottom right corners of the front and the back face. xtl1, ytl1, xbr1, ybr1, xtl2, ytl2, xbr2, ybr2

Ordered bounding box

For ordered bounding boxes, the point_order property is set to a value between 1 and 8. It can be interpreted as follows:

Value
Description

-1

Bounding box was created without point order.

1

The top-left corner of an un-rotated bounding-box is the first point. The other points are in clockwise direction.

2

The top-right corner of an un-rotated bounding-box is the first point. The other points are in clockwise direction.

3

The bottom-right corner of an un-rotated bounding-box is the first point. The other points are in clockwise direction.

4

The bottom-left corner of an un-rotated bounding-box is the first point. The other points are in clockwise direction.

5

The top-left corner of an un-rotated bounding-box is the first point. The other points are in anti-clockwise direction.

6

The top-right corner of an un-rotated bounding-box is the first point. The other points are in anti-clockwise direction.

7

The bottom-right corner of an un-rotated bounding-box is the first point. The other points are in anti-clockwise direction.

8

The bottom-left corner of an un-rotated bounding-box is the first point. The other points are in anti-clockwise direction.

Tracks

Tracks are tracked objects across frames with the same ID. The same object across different frames has different annotation IDs but the same track_id.

The property occluded specifies if the object is part of the same track but not visible in the current frame.

Additionally the related_track_ids describes an array of other tracks that are connected with a particular tracked object.

Last updated