JSON file for training

[kzchua1998]

Hi, first of all thank you so much for your source codes, actually I am curious about how should the json file be created for training. Let say I have 500 images for training, does each image associate with a standalone json file, or should all the annotations for the images share a single json file, if so how should they be arranged? Thank you in advance.

[duanzhiihao]

Hi, thank you for your interest.

The short answer is that all the annotations share a single JSON file. If you know the COCO dataset, we use the same annotation format as theirs, except that for each bounding box there is an angle in our case.

Below is an example JSON file. This dataset contains only one image.
https://github.com/duanzhiihao/RAPiD/blob/master/images/tiny_val/one.json

On the top level, the JSON file should have three keys, each maps to a list:

{
 "annotations": [ ... ],
 "images": [ ... ],
 "categories": [ ... ]
}

Below is an example of what each list looks like:

{
 "annotations": [
  {
   "area": 13973.992515539998, # area of the bounding box
   "bbox": [
    510.4806, # center x
    468.7095, # center y
    100.6869, # width (should be smaller than height)
    138.7866, # height
    -17.188751454018586 # angle (degree, clock-wsie)
   ],
   "category_id": 1, # ID for object class
   "image_name": "ytb_000001.jpg", # file name of the image
   "image_id": "ytb_000001", # ID of the image
   "iscrowd": 0, # A dummy key. Just let it be zero.
   "segmentation": [] # A dummy key. Just let it be an empty list.
  },
 ],
 "images": [
  {
   "file_name": "ytb_000001.jpg", # file name to find the image
   "id": "ytb_000001", # unique ID of the image
   "width": 1015, # image width
   "height": 1010 # image height
  }
 ],
 "categories": [
  {
   "id": 1, # ID for object class
   "name": "person", # human understandable name of this calss
   "supercategory": "person" # A dummy key. Useless in our code.
  }
 ]
}

Please also check our CEPDOF dataset for details of the annotation format.

[kzchua1998]

I see, that really helps, thank you for the quick reply. Have a nice day.