Dictionary schema for datasets

[broadinstitute]

Dictionary schema for datasets

[jhung0]

COCO is too complicated.

We want to be able to include bounding boxes and/or masks (segmentations). Also, we want the structure to be flexible to additional data like metadata, class hierarchies or multilabel, difficult classes.

[jhung0]

For the option of class hierarchies, maybe let people include tree information

[0x00b1]

Here’s my current thinking:

{
  "pathname": "image.png",
  "objects": [
    {
      "class": "foo",
      "mask": "mask.png",
      "x_maximum": 1,
      "x_minimum": 1,
      "y_maximum": 1,
      "y_minimum": 1
    }
  ]
}

I like this too:

{
  "pathname": "image.png",
  "objects": [
    {
      "class": "foo",
      "mask": "mask.png",
      "maximum": [1, 1],
      "minimum": [1, 1]
    }
  ]
}

"mask" is a pathname to an 8-bit mask where 255 values identify the object.