json ” coco = COCO ( ann_file ) # Get list of category_ids, here for bicycleĬategory_ids = coco. Then we can use the COCO api to get a list of all image_ids which contain annotated bicycles.

To get annotated bicycle images we can subsample the COCO dataset for the bicycle class (coco label 2).įirst, we clone the repository and add the folders images and annotations to the root of the repository. If you don’t want to write your own code to access the annotations you can get the COCO api.Īs a brief example let’s say we want to train a bicycle detector. The bounding box field provides the bounding box coordinates in the COCO format x,y,w,h where (x,y) are the coordinates of the top left corner of the box and (w,h) the width and height of the box. Each annotation is uniquely identifiable by its id ( annotation_id). The image_id maps this annotation to the image object, while the category_id provides the class information. These map the category_id to the classname. The categories key contains a list of category objects. Image name without file extension and leading zeros For example, the first license is īefore we explore the remaining objects let’s define three variables. and the licenses under which the images are used. The first two objects contain information regarding the dataset such as date of creation etc. import json path = './annotations/instances_train2017.json' f = open ( path ) anns = json. In Python we can access these objects by loading the file with the json module.

Downloading Coco