-
Reads the annotations from a JSON file (i.e.
"instances_train2017.json"
) -
Reads the images either from the disk (i.e. from
"COCO/train2017/"
folder), or by downloading them fromcoco_urls
provided in annotations JSON file -
Selects all of the Coco classes (
81
classes total), or produces a subset to scale according to the project (i.e. ifclasses = ['person']
, it will find all of the images with a person, and delete all other bounding boxes and masks from them) -
Pads the images so their shapes are a multiple of some number (i.e. a multiple of
32
for aVGG-FCN
architecture) -
Shows the image alongside the semantic masks of the objects
-
Fits easily into a
pytorch
DataLoader (see Usage_Coco_Manager.ipynb)
Coco_Manager.ipynb: for being imported in a project
Usage_Coco_Manager.ipynb: use case and functionality explanation