{
  "info" : info,
  "images" : [image],
  "categories" : [category],
  "annotations" : [annotation]
}

info 
{
  # Required
  "version" : str,
  "description" : str,
  
  # Optional
  "year" : int,
  "contributor" : str
  "date_created" : datetime
}

image
{
  # Required
  "id" : str,
  "width" : int,
  "height" : int,
  "file_name" : str,
  
  # Optional
  "rights_holder" : str,    
  "datetime": datetime,  
  "seq_id": str,
  "seq_num_frames": int,
  "frame_num": int
  
  # This is an int in older data sets, but convention is now strings
  "location": str,
  
  # Image corruption is quite common in camera trap images, and throwing out corrupt
  # images in database assembly is "dodging part of the problem".  Wherever possible,
  # use this flag to indicate that an image failed to load, e.g. in PIL and/or TensorFlow.
  "corrupt": bool
}

category
{
  # Required
  
  # Category ID 0 reserved for the class "empty"; all other categories vary by data
  # set.  Positive integers only.
  "id" : int,
  "name" : str  
}

annotation
{
  # Required
  "id" : str,
  "image_id" : str,  
  "category_id" : int,
  
  # Optional
  
  # These are in absolute, floating-point coordinates, with the origin at the upper-left
  "bbox": [x,y,width,height],
  
  # This indicates that this annotation is really applied at the *sequence* level,
  # and may not be reliable at the individual-image level.  Since the *sequences* are
  # the "atom of interest" for most ecology applications, this is common.
  "sequence_level_annotation" : bool
}