Transforming COCO datasets

COCO is a format for specifying large-scale object detection, segmentation, and captioning datasets. This Python example shows you how to transform a COCO object detection format dataset into an Amazon Rekognition Custom Labels bounding box format manifest file. This section also includes information that you can use to write your own code.

A COCO format JSON file consists of five sections providing information for an entire dataset. For more information, see COCO format.

• info – general information about the dataset.

• licenses – license information for the images in the dataset.

• images – a list of images in the dataset.

• annotations – a list of annotations (including bounding boxes) that are present in all images in the dataset.

• categories – a list of label categories.

You need information from the images, annotations, and categories lists to create an Amazon Rekognition Custom Labels manifest file.

An Amazon Rekognition Custom Labels manifest file is in JSON lines format where each line has the bounding box and label information for one or more objects on an image. For more information, see Object localization in manifest files.

Mapping COCO Objects to a Custom Labels JSON Line

To transform a COCO format dataset, you map the COCO dataset to an Amazon Rekognition Custom Labels manifest file for object localization. For more information, see Object localization in manifest files. To build a JSON line for each image, the manifest file needs to map the COCO dataset image, annotation, and category object field IDs.

The following is an example COCO manifest file. For more information, see COCO format.

{
    "info": {
        "description": "COCO 2017 Dataset","url": "http://cocodataset.org","version": "1.0","year": 2017,"contributor": "COCO Consortium","date_created": "2017/09/01"
    },
    "licenses": [
        {"url": "http://creativecommons.org/licenses/by/2.0/","id": 4,"name": "Attribution License"}
    ],
    "images": [
        {"id": 242287, "license": 4, "coco_url": "http://images.cocodataset.org/val2017/xxxxxxxxxxxx.jpg", "flickr_url": "http://farm3.staticflickr.com/2626/xxxxxxxxxxxx.jpg", "width": 426, "height": 640, "file_name": "xxxxxxxxx.jpg", "date_captured": "2013-11-15 02:41:42"},
        {"id": 245915, "license": 4, "coco_url": "http://images.cocodataset.org/val2017/nnnnnnnnnnnn.jpg", "flickr_url": "http://farm1.staticflickr.com/88/xxxxxxxxxxxx.jpg", "width": 640, "height": 480, "file_name": "nnnnnnnnnn.jpg", "date_captured": "2013-11-18 02:53:27"}
    ],
    "annotations": [
        {"id": 125686, "category_id": 0, "iscrowd": 0, "segmentation": [[164.81, 417.51,......167.55, 410.64]], "image_id": 242287, "area": 42061.80340000001, "bbox": [19.23, 383.18, 314.5, 244.46]},
        {"id": 1409619, "category_id": 0, "iscrowd": 0, "segmentation": [[376.81, 238.8,........382.74, 241.17]], "image_id": 245915, "area": 3556.2197000000015, "bbox": [399, 251, 155, 101]},
        {"id": 1410165, "category_id": 1, "iscrowd": 0, "segmentation": [[486.34, 239.01,..........495.95, 244.39]], "image_id": 245915, "area": 1775.8932499999994, "bbox": [86, 65, 220, 334]}
    ],
    "categories": [
        {"supercategory": "speaker","id": 0,"name": "echo"},
        {"supercategory": "speaker","id": 1,"name": "echo dot"}
    ]
}

The following diagram shows how the COCO dataset lists for a dataset map to Amazon Rekognition Custom Labels JSON lines for an image. Matching colors indicate information for a single image.

To get the COCO objects for a single JSON line

1. For each image in the images list, get the annotation from the annotations list where the value of the annotation field image_id matches the image id field.

2. For each annotation matched in step 1, read through the categories list and get each category where the value of the category field id matches the annotation object category_id field.

3. Create a JSON line for the image using the matched image, annotation, and category objects. To map the fields, see Mapping COCO object fields to a Custom Labels JSON line object fields.

4. Repeat steps 1–3 until you have created JSON lines for each image object in the images list.

For example code, see Transforming a COCO dataset.

Mapping COCO object fields to a Custom Labels JSON line object fields

After you identify the COCO objects for an Amazon Rekognition Custom Labels JSON line, you need to map the COCO object fields to the respective Amazon Rekognition Custom Labels JSON line object fields. The following example Amazon Rekognition Custom Labels JSON line maps one image (id=000000245915) to the preceding COCO JSON example. Note the following information.

• source-ref is the location of the image in an Amazon S3 bucket. If your COCO images aren't stored in an Amazon S3 bucket, you need to move them to an Amazon S3 bucket.

• The annotations list contains an annotation object for each object on the image. An annotation object includes bounding box information (top, left,width, height) and a label identifier (class_id).

• The label identifier (class_id) maps to the class-map list in the metadata. It lists the labels used on the image.

{
	"source-ref": "s3://custom-labels-bucket/images/000000245915.jpg",
	"bounding-box": {
		"image_size": {
			"width": 640,
			"height": 480,
			"depth": 3
		},
		"annotations": [{
			"class_id": 0,
			"top": 251,
			"left": 399,
			"width": 155,
			"height": 101
		}, {
			"class_id": 1,
			"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": "2018-10-18T22:18:13.527256",
		"job-name": "my job"
	}
}

Use the following information to map Amazon Rekognition Custom Labels manifest file fields to COCO dataset JSON fields.

source-ref

The S3 format URL for the location of the image. The image must be stored in an S3 bucket. For more information, see source-ref. If the coco_url COCO field points to an S3 bucket location, you can use the value of coco_url for the value of source-ref. Alternatively, you can map source-ref to the file_name (COCO) field and in your transform code, add the required S3 path to where the image is stored.

bounding-box

A label attribute name of your choosing. For more information, see bounding-box.

image_size

The size of the image in pixels. Maps to an image object in the images list.

• height-> image.height

• width-> image.width

• depth-> Not used by Amazon Rekognition Custom Labels but a value must be supplied.

annotations

A list of annotation objects. There’s one annotation for each object on the image.

annotation

Contains bounding box information for one instance of an object on the image.

• class_id -> numerical id mapping to Custom Label’s class-map list.

• top -> bbox[1]

• left -> bbox[0]

• width -> bbox[2]

• height -> bbox[3]

bounding-box-metadata

Metadata for the label attribute. Includes the labels and label identifiers. For more information, see bounding-box-metadata.

Objects

An array of objects in the image. Maps to the annotations list by index.

Object

• confidence->Not used by Amazon Rekognition Custom Labels, but a value (1) is required.

class-map

A map of the labels (classes) that apply to objects detected in the image. Maps to category objects in the categories list.

• id -> category.id

• id value -> category.name

type

Must be groundtruth/object-detection

human-annotated

Specify yes or no. For more information, see bounding-box-metadata.

creation-date -> image.date_captured

The creation date and time of the image. Maps to the image.date_captured field of an image in the COCO images list. Amazon Rekognition Custom Labels expects the format of creation-date to be Y-M-DTH:M:S.

job-name

A job name of your choosing.