What is a Ground Truth file?

Ground Truth is a term that came from meteorology, which refers to information collected from a single location. When we look at this from a machine learning perspective, Ground Truth is used for showing the source or actual value of the data that want to learn.

Ground Truth is a term that refers to real-world data used to learn and test AI model outputs in the field of artificial intelligence data. Ground Truth data is necessary for a lot of AI applications such as autonomous driving and audio or voice recognition.

The GT file upload format only supports JSON.
Upload conditions: It must have the same folder structure and file name as the uploaded resource file, in the JSON file structure.

The task type set in the project (box, keypoint, etc.), class, and attribute names and values must be the same.
Even if the case and spacing are different, the upload fails.
Example of GT file upload failure (1)
Class name registered in the project: car /
The class name is written in JSON: Car
Failure case due to conditions not in the project (2)
The value of the attribute ‘weather’ under class A registered in the project: sunny, cloudy, rainy
The class name is written in JSON: Snow

GT upload format and validity

GT Upload Schema

{
  "annotations": [
    {
      "id": "26-abd21cb9-cb89-4cf4-8de6-b0a6421024b3",
      "type": "poly_seg",
      "attributes": {},
      "points": [
        [
          905,
          44
        ],
        [
          1280,
          228
        ],
        [
          1272,
          232
        ]
      ],
        "label": "Road"
      }
    ],
    "attributes": {},
    "file_id": "6465703408f21776587bbdac",
    "filename": "1020312_0_2023-05-03_17-10-44_1683133844_0001584.jpg",
    "parent_path": "/0000/3576612_0_2023-04-03_17-10-44_1683133844",
    "metadata": {
      "height": 1080,
      "width": 1920
    }
  }

GT Upload Field Definition

Name	Type	Description	Example
annotations	list[object]	Annotation information	Reference annotation field
attributes	object	When setting an attribute of the task type classification, it is the same as the attribute of the classification When there is no classification {}	{ "attribute name”: “attribute value”, "attribute name”: “attribute value" } When it isn’t set ⇒ {}
label	string	It exists only when the task type is the classification *If there is no classification in task type, the field does not exist	“Classification”
file_id	object	The ID of the file	620aea6c1e3c3ccd9cfd8bc9
scene_id	string	The ID of the scene file * If the resource type is neither Sequence Image nor Point Cloud, the field does not exist	620aea6c1e3c3ccd9cfd8bc8
filename	string	File names * In the case of the resource type Sequence Image or Point Cloud, this corresponds to a frame within the scene file	image (28).jpeg
parent_path	string	The parent path of the file	/batch_5/20220624/daytime
last_modifier_id	string	Last modifier id	5fc7323e8511afbb60e8554b
metadata	object	File metadata by each resource type Image: image size PointCloud, Text: {} empty value Video:fps (frames)	# If it's an image{ "height": 1080, "width": 1920 } # If it's a video{ “fps”: 30 }
last_workers	object	The last worker at each stage	{ "labeling": "5fc7323e8511afbb60e8554b", "inspection": null, "final inspection": null }
tags	list[object]	The tag value of the file *If you do not check whether the tag is included or not, the field does not exist	When the tag is applied ⇒ [”car”] When the tag is not applied ⇒ no field
parent_tags	list[object]	The tag value of the folder *If you do not check whether the tag is included or not, the field does not exist	If the tag is applied ⇒ [” Primary task”] If the tag is not applied ⇒ no field

GT file generation guide for each field

Annotations field
1. ID definition
  1. Starts with number, and the string part is written in uuid format.
  - ex. 26-abd21cb9-cb89-4cf4-8de6-b0a6421024b3
2. Attributes definition
  1. Key value and value must be registered to labeling setting.
3. Type definition
  1. [Single/Sequence Image(Vector)] : In case of Bounding Box, Polygon, Polyline, Segmentation
    - annotations > points: Co-ordinate are to be real number or integer value and must not deviate from working range.
Files matching with parent_path and filename must not be started.
- However, both parent_path and filename must be written.
Attributes field
1. Key value and value must be registered to labeling setting.
Annotation track_id must not be duplicated in sequence_image.
Upload will not be effected if file_id information is omitted.

Validation

Status validation: Check the status that the task has not started.
Check for the existence of file paths: Check that the actual file exists in the path name.
Annotation check: Check the annotation data that matches your project settings.

Refer to AIMMO enterprise output form

If you have any other inquiries, please get in touch with us at [email protected]