Object localization in manifest files

You can import images labeled with object localization information by adding SageMaker Ground Truth Bounding Box Job Output format JSON lines to a manifest file.

Localization information represents the location of an object on an image. The location is represented by a bounding box that surrounds the object. The bounding box structure contains the upper-left coordinates of the bounding box and the bounding box's width and height. A bounding box format JSON line includes bounding boxes for the locations of one or more an objects on an image and the class of each object on the image.

A manifest file is made of one or more JSON lines, each line contains the information for a single image.

To create a manifest file for object localization

1. Create an empty text file.

2. Add a JSON line for each image the that you want to import. Each JSON line should look similar to the following.

   {"source-ref": "s3://bucket/images/IMG_1186.png", "bounding-box": {"image_size": [{"width": 640, "height": 480, "depth": 3}], "annotations": [{ "class_id": 1,	"top": 251,	"left": 399, "width": 155, "height": 101}, {"class_id": 0, "top": 65, "left": 86, "width": 220,	"height": 334}]}, "bounding-box-metadata": {"objects": [{ "confidence": 1}, {"confidence": 1}],	"class-map": {"0": "Echo",	"1": "Echo Dot"}, "type": "groundtruth/object-detection", "human-annotated": "yes",	"creation-date": "2013-11-18T02:53:27", "job-name": "my job"}}

3. Save the file. You can use the extension .manifest, but it is not required.

4. Create a dataset using the file that you just created. For more information, see To create a dataset using a SageMaker Ground Truth format manifest file (console).

Object bounding Box JSON lines

In this section, we show you how to create a JSON line for a single image. The following image shows bounding boxes around Amazon Echo and Amazon Echo Dot devices.

The following is the bounding box JSON line for the preceding image.

{
	"source-ref": "s3://custom-labels-bucket/images/IMG_1186.png",
	"bounding-box": {
		"image_size": [{
			"width": 640,
			"height": 480,
			"depth": 3
		}],
		"annotations": [{
			"class_id": 1,
			"top": 251,
			"left": 399,
			"width": 155,
			"height": 101
		}, {
			"class_id": 0,
			"top": 65,
			"left": 86,
			"width": 220,
			"height": 334
		}]
	},
	"bounding-box-metadata": {
		"objects": [{
			"confidence": 1
		}, {
			"confidence": 1
		}],
		"class-map": {
			"0": "Echo",
			"1": "Echo Dot"
		},
		"type": "groundtruth/object-detection",
		"human-annotated": "yes",
		"creation-date": "2013-11-18T02:53:27",
		"job-name": "my job"
	}
}

Note the following information.

source-ref

(Required) The Amazon S3 location of the image. The format is "s3://BUCKET/OBJECT_PATH". Images in an imported dataset must be stored in the same Amazon S3 bucket.

bounding-box

(Required) The label attribute. You choose the field name. Contains the image size and the bounding boxes for each object detected in the image. There must be corresponding metadata identified by the field name with -metadata appended. For example, "bounding-box-metadata".

image_size

(Required) A single element array containing the size of the image in pixels.

• height – (Required) The height of the image in pixels.

• width – (Required) The depth of the image in pixels.

• depth – (Required) The number of channels in the image. For RGB images, the value is 3. Not currently used by Amazon Rekognition Custom Labels, but a value is required.

annotations

(Required) An array of bounding box information for each object detected in the image.

• class_id – (Required) Maps to the label in class-map. In the preceding example, the object with the class_id of 1 is the Echo Dot in the image.

• top – (Required) The distance from the top of the image to the top of the bounding box, in pixels.

• left – (Required) The distance from the left of the image to the left of the bounding box, in pixels.

• width – (Required) The width of the bounding box, in pixels.

• height – (Required) The height of the bounding box, in pixels.

bounding-box-metadata

(Required) Metadata about the label attribute. The field name must be the same as the label attribute with -metadata appended. An array of bounding box information for each object detected in the image.

Objects

(Required) An array of objects that are in the image. Maps to the annotations array by index. The confidence attribute isn't used by Amazon Rekognition Custom Labels.

class-map

(Required) A map of the classes that apply to objects detected in the image.

type

(Required) The type of classification job. "groundtruth/object-detection" identifies the job as object detection.

creation-date

(Required) The Coordinated Universal Time (UTC) date and time that the label was created.

human-annotated

(Required) Specify "yes", if the annotation was completed by a human. Otherwise "no".

job-name

(Optional) The name of the job that processes the image.