jsk1107 / coco_utils

Split train/val/test coco dataset and Adjust annotations & categories.

coco_utils

When pre-processing and post-processing image, sometimes you need to add, adjust, or remove categories in the coco json file. Also, train, val, and test should be separated from one json file. coco_utils is a module that provides these functions.

Using this, you will be able to process coco json files easily.

Install

You only need to install the pycocotools.

$ git clone https://github.com/jsk1107/coco_utils.git
$ pip install pycocotools

Usage

from coco import CocoUtils

PATH = '{your_coco_json_file}'
coco_util = CocoUtils(PATH)

import json
from coco import CocoUtils

with open(PATH, 'r') as f:
    coco = json.load(f)
coco_util = CocoUtils(coco)

• Even if you have already opened a COCO Json file, you can create a class instance.
• Of course, all functions provided by pycocotools are also available.

delete category

coco_util.del_category("person")
# or
coco_util.del_category(["person", "dog", "book", ...])

• Enter an argument of type str or list.
• All information of categories and annotations is lost.
• If it is an wrong category name, nothing is deleted.

adjust category

coco_util.adj_category(bf_catNms="person", af_catNms="people")
# or
coco_util.del_category(bf_catNms=["person", "dog"], af_catNms=["people", "fox"]) # ok
coco_util.del_category(bf_catNms=["people", "dogs"], af_catNms=["human", "fox"]) # error. because people dose not exist.

• Enter an argument of type str or list.
• Match ordered pairs of arguments.
• If you adjust a category(bf_catNms) that does not exist, an error occurs.
• About category_id of annotations, category_id corresponding to bf_Nms is changed to category_id corresponding to af_Nms.

add category

coco_util.add_category('hero')
# or
coco_util.add_category(['hero', 'iron man'])

• Enter an argument of type str or list.
• If you enter a category that already exists, an error occurs.
• Only added to categories. annotation information cannot be added.

split train/val/test

coco_util.split_train_val_test(val_ratio=.2, test_ratio=.1, save_dir='{your_save_dir}', set_seed=1)

• If test_ratio is not none, two json files instance_default_train.json and instance_default_val.json are created.
• If you set the seed, you can get the same split result.