Skip to content

Latest commit

 

History

History
614 lines (575 loc) · 12.1 KB

import-annotations.md

File metadata and controls

614 lines (575 loc) · 12.1 KB
Raw

Import existing annotations / predictions

Create an annotation folder as such:

data-test   
│
│───images
│   │    xxx.jpg
│   └─── yyy.jpg
│       
└───annotations
    │─── task1.json
    └─── task1
        │    xxx.json
        └─── yyy.json

The task1.json file contains global task settings (task type, task category, image folder, etc) and its correspoding task1 folder contains an annotation file for each image.

Common import information

For each task you must add a corresponding task json file with the following structure:

// task1.json
{
 "name": "task1",
 "spec": {
  "plugin_name": "rectangle", // or "polygon", "segmentation", "keypoints", "cuboid", "segmentation-interactive", "smart-rectangle", etc
  "label_schema": {
   "category": [
    {
     "name": "car",
     "color": "green",
     "properties": [ // optional additional properties for given category (its color, its brand, its subtype, etc), see note below
      {
       "name": "isBlue",
       "type": "checkbox",
       "default": false
      }
     ]
    },
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "size",
       "type": "dropdown",
       "enum": [
        "little",
        "big"
       ],
       "default": "little"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "images/", // relative path of the image folder
  "data_type": "image" // type of the data: "image" or "pcl" or "sequence_image" or "sequence_pcl"
 }
}

Note: for each possible object category, you can list additionnal properties which are either:

  • checkbox (boolean value) : default value is true or false (make sure not to add quotation marks)
  • dropdown type (list of choices) : default value if one of the choices. We advise you to select default category and properties as the most frequent ones.

Note 2: When using sequence data type, make sure to prefix the plugin name by sequence- (eg. sequence-polygon), except for specific plugins which only work with sequences (eg. tracking).

Note 3: Unless specified otherwise, all 2d object coordinates are normalized by the image dimension. The object lines drawn in-between pixels, 0 corresponds to the left (or top) of the pixel 0, and 1 corresponds to the right (or bottom) of the right-most (or bottom-most) pixel.

Import bounding box annotations / predictions

box

Example of annotation specification file for bounding box annotation.

// task1.json
{
 "name": "task1",
 "spec": {
  "plugin_name": "rectangle",
  "label_schema": {
   "category": [
    {
     "name": "car",
     "color": "green",
     "properties": [
      {
       "name": "isOccluded",
       "type": "checkbox",
       "default": false
      }
     ]
    },
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "age",
       "type": "dropdown",
       "enum": [
        "child",
        "adult",
        "elderly"
       ],
       "default": "adult"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "images/",
  "data_type": "image"
 }
}

Example of annotation file for bounding box annotation:

// xxx.json
{
 "task_name": "task1",
 "annotations": [
  {
   "id": "m83ihfsoplq",
   "geometry": {
    "vertices": [
     0.4001610305958132, // left
     0.23466666666666666, // top
     0.5048309178743962, // right
     0.6906666666666667 // bottom
    ],
    "type": "rectangle"
   },
   "category": "person",
   "options": {
    "size": "child"
   }
  }
 ],
 "data": {
  "type": "image",
  "children": "",
  "path": "images/xxx.jpg"
 }
}

Import polygon annotations / predictions

polygon

Example of annotation specification file for polygon annotation.

// task1.json
{
 "name": "task2",
 "spec": {
  "plugin_name": "polygon",
  "label_schema": {
   "category": [
    {
     "name": "car",
     "color": "green",
     "properties": [
      {
       "name": "isOccluded",
       "type": "checkbox",
       "default": false
      }
     ]
    },
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "age",
       "type": "dropdown",
       "enum": [
        "child",
        "adult",
        "elderly"
       ],
       "default": "adult"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "images/",
  "data_type": "image"
 }
}

Example of annotation file for polygon annotation:

// xxx.json
{
 "task_name": "task2",
 "annotations": [
  {
   "id": "m83ihfsoplq",
   "geometry": {
    "vertices": [
     0.4001610305958132, // pt1 x
     0.23466666666666666, // pt1 y
     0.5048309178743962, // pt2 x
     0.6906666666666667, // pt2 y
     0.3048309178743962, // pt3 x
     0.8906666666666667 // pt3 y
    ],
    "type": "polygon"
   },
   "category": "person",
   "options": {
    "size": "child"
   }
  },
  {
   "id": "pmgfqzs2i9k0",
   "geometry": {
    "mvertices": [
     [0.5587761674718197,0.296,0.49033816425120774,0.696,0.6900161030595813,0.7973333333333333,0.715780998389694,0.36533333333333334],
     [0.2898550724637681,0.3253333333333333,0.20933977455716588,0.6186666666666667,0.3325281803542673,0.8586666666666667,0.42673107890499196,0.4053333333333333]
    ]
    "vertices":[],
    "type":"multi_polygon"
   },
   "category": "person",
   "options": {
    "size": "child"
   }
  }
 ],
 "data": {
  "type": "image",
  "children": "",
  "path": "images/xxx.jpg"
 }
}

Note: notice that occluded objects can be formatted as multiple associated polygons : "mvertices".

Import keypoints annotations / predictions

keypoints

Example of annotation specification file for keypoints annotation.

// task1.json
{
 "name": "task3",
 "spec": {
  "plugin_name": "keypoints",
  "label_schema": {
   "category": [
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "age",
       "type": "dropdown",
       "enum": [
        "child",
        "adult",
        "elderly"
       ],
       "default": "adult"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "images/",
  "data_type": "image"
 }
}

Example of annotation file for skeleton annotation:

// xxx.json
{
 "task_name": "task3",
 "annotations": [
  {
    "id":"4nuyc8i1gbs",
    "geometry":{
        "vertices":[0.7093750000000001,0.40416666666666673,0.5890625,0.17708333333333337,0.696875,0.2], // example of skeleton made of 3 points (x,y,x,y,x,y)
        "edges":[[0,1],[0,2]], // edges to be drawn between the keypoints
        "visibles":[true,true,true], // whether each keypoint is visible or not
        "type":"graph"
     },
     "category":"person",
     "options":{
        "size": "adult"
     }
   }
 ],
 "data": {
  "type": "image",
  "children": "",
  "path": "images/xxx.jpg"
 }
}

Import segmentation annotation / prediction

segmentation

Example of annotation specification file for panoptic segmentation annotation.

// task1.json
{
 "name": "task4",
 "spec": {
  "plugin_name": "segmentation",
  "label_schema": {
      "category": [
        {
          "name": "car",
          "color": "green",
          "idx": 1,
          "instance": true
        },
        {
          "name": "road",
          "color": "blue",
          "idx": 2,
          "instance": false
        }
      ],
      "default": "car"
    }
 },
 "dataset": {
  "path": "images/",
  "data_type": "image"
 }
}

Note that for each category you must specify if it is a semantic or instance category. Also note that the optional additional properties are not supported for the segmentation plugin.

Example of annotation file for segmentation annotation:

// xxx.json
{
 "task_name": "task4",
 "annotations": [
  {
   "id": 0,
   "mask": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgA...ASUVORK5CYII="
  }
 ],
 "data": {
  "type": "image",
  "children": "",
  "path": "images/xxx.jpg"
 }
}

The segmentation mask is base64 encoded and can be read and written using the following code:

import json
import base64
import cv2
import numpy as np

def readb64(uri):
  encoded_data = uri.split(',')[1]
  nparr = np.fromstring(base64.b64decode(encoded_data), np.uint8)
  img = cv2.imdecode(nparr, cv2.IMREAD_UNCHANGED)
  return img

def writeb64(img):
    retval, buffer = cv2.imencode('.png', img)
    pic_str = base64.b64encode(buffer)
    pic_str = pic_str.decode()
    return pic_str

# assuming you stored the mask in a json file of the following structure
# { annotations: [{mask: "data:image/png;base64,iVBORw0KGgoAA..."}]}
with open(filename, 'r') as f:
  annotations = json.load(f)["annotations"]
for ann in annotations:
  mask = readb64(ann["mask"])
  print(mask.shape)
  # should be (height,width,3)
  # corresponding to [id1, id2, classIdx]
  # id1 and id2 are zeros in case of semantic segmentation
  # in case of instance segmentation, they should be incremental: (id1:1, id2:0), then (id1:2, id2:0), then (id1:3, id2:0), etc...

Import bounding box tracking annotations / predictions

Example of annotation specification file for bounding box tracking annotation.

// task1.json
{
 "name": "task5",
 "spec": {
  "plugin_name": "tracking",
  "label_schema": {
   "category": [
    {
     "name": "car",
     "color": "green",
     "properties": [
      {
       "name": "isOccluded",
       "type": "checkbox",
       "default": false
      }
     ]
    },
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "age",
       "type": "dropdown",
       "enum": [
        "child",
        "adult",
        "elderly"
       ],
       "default": "adult"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "images/",
  "data_type": "image"
 }
}

Example of annotation file for bounding box tracking annotation:

// xxx.json
{
 "task_name": "task5",
 "annotations": {
  "0": {
   "id": "0",
   "keyShapes": {
    "0": {
     "geometry": {
      "vertices": [
       0.24798711755233493,
       0.3413333333333333,
       0.4500805152979066,
       0.7093333333333334
      ],
      "type": "rectangle"
     },
     "timestamp": 0,
     "labels": {}
    },
    "2": {
     "geometry": {
      "vertices": [
       0.24798711755233493,
       0.26458333333333334,
       0.8203125,
       0.7093333333333334
      ],
      "type": "rectangle"
     },
     "timestamp": 2,
     "labels": {},
     "id": "0",
     "color": "#ff1100"
    }
   },
   "category": "car",
   "labels": {}
  }
 },
 "data": {
  "type": "sequence_image",
  "children": [
   {
    "timestamp": 0,
    "path": "images/vid1/000.png"
   },
   {
    "timestamp": 1,
    "path": "images/vid1/001.png"
   },
   {
    "timestamp": 2,
    "path": "images/vid1/002.jpg"
   }
  ],
  "path": "images/vid1/"
 }
}

Import cuboid annotations / predictions

cuboid

Example of annotation specification file for cuboid annotation.

// task1.json
{
 "name": "task6",
 "spec": {
  "plugin_name": "cuboid",
  "label_schema": {
   "category": [
    {
     "name": "car",
     "color": "green",
     "properties": [
      {
       "name": "isOccluded",
       "type": "checkbox",
       "default": false
      }
     ]
    },
    {
     "name": "person",
     "color": "#eca0a0",
     "properties": [
      {
       "name": "age",
       "type": "dropdown",
       "enum": [
        "child",
        "adult",
        "elderly"
       ],
       "default": "adult"
      }
     ]
    }
   ],
   "default": "person"
  }
 },
 "dataset": {
  "path": "pcls/",
  "data_type": "pcl"
 }
}

Example of annotation file for bounding box annotation:

// xxx.json
{
 "task_name": "task6",
 "annotations": [
  {
   "position": [
    0.4802595234459003, // x
    -0.06460804884089505, // y
    -0.45750510778452735 // z
   ],
   "size": [
    3.0867916079091895, // length
    1.29858273838599, // width
    0.9156043512339238 // height
   ],
   "heading": 0.03437769738776719,
   "id": "0.7u894bwwk49",
   "category": "car",
   "options": {
    "brand": "fiat"
   }
  }
 ],
 "data": {
  "type": "pcl",
  "children": "",
  "path": "pcls/xxx.bin"
 }
}