readme.md

Annotation data formats

There are several data annotation formats allowed for TensorBox. The most simple is json-file format which presents the data structure well.

Original TensorBox

Original TensorBox expects the list of objects each of which describes one annotations for one image.

[image]
  
image{
  "image_path": string,
  "rects": [rect]
}
  
rect{
  "x1": int,
  "y1": int,
  "x2": int,
  "x2": int
}

Each annotation is an object with two properties: image_path (string) and rects (list). The second property describes all bounding boxes which present on the current image. The format of bounding box description consists of four integer properties which mean the main diagonal of the rectangle (x1, y1) - (x2,y2). TensorBox reading procedure expects that x1<x2 and y1<y2. Example:

[
  {
    "image_path": "images/1/abc.jpg",
    "rects":
      [
        {"x1": 0, "y1": 0, "x2": 100, "y2": 100},
        {"x1": 200, "y1": 150, "x2": 220, "y2": 300}
      ]
  },
  {
    "image_path": "images/2/klm.jpg",
    "rects":
      [
        {"x1": 200, "y1": 0, "x2": 300, "y2": 100}
      ]
  },
]

This fork extension

In this fork we suggest the following extension:

{
"images": [image],
"classes": [string]
}
  
image{
  "id": int,
  "image_path": string,
  "rects": [rect]
}
  
rect{
  "classID": int,
  "x1": int,
  "y1": int,
  "x2": int,
  "x2": int
}

This extension allows to point class of object which is surrounded by each box. The classID value is the index in classes collection. This extension is necessary step towards to the multiclass TensorBox model which we hope will be implemented later.