cocojson

Utility functions for COCO json annotation format. The COCO Format is defined here. For COCO Evaluation, please use these instead: Official or my fork.

Install

cocojson is available on pypi through pip3 install cocojson
or if you prefer, clone this repo and it can be installed through pip3 install -e . (editable install) or pip3 install . as well.

Organisation

Codes to "Utility Tools" can be found in ./cocojson/tools.
Codes to "Converters" can be found in ./cocojson/convert.

Usage

Please click into each for more details (if applicable). Links works only if you're viewing from the github homepage.

Utility Tools

COCO Categori-fy

Convert your custom dataset into COCO categories. Usually used for testing a coco-pretrained model against a custom dataset with overlapping categories with the 80 COCO classes.

python3 -m cocojson.run.coco_catify -h

Extract only Annotations

Get annotations/predictions only from a COCO JSON. Usually used to generate a list of predictions for COCO evaluation.

python3 -m cocojson.run.pred_only -h

Filter Categories

Filter categories from COCO JSON.

python3 -m cocojson.run.filter_cat -h

Insert Images Meta-Information

Insert any extra attributes/image meta information associated with the images into the coco json file.

python3 -m cocojson.run.insert_img_meta -h

Map Categories

Mapping categories to a new dataset. Usually used for converting annotation labels to actual class label for training.

python3 -m cocojson.run.map_cat -h

Match Images between 2 COCO JSONs

Match images between a reference COCO JSON A and COCO JSON B (to be trimmed). Any images in JSON B that is not found in JSON A will be removed (along with associated annotations)

python3 -m cocojson.run.match_imgs -h

Merge

Merges multiple datasets

python3 -m cocojson.run.merge -h

Merge from file

Merges multiple datasets

python3 -m cocojson.run.merge_from_file -h

Merge JSONs

Merges multiple coco jsons.

python3 -m cocojson.run.merge_jsons -h

Prune Ignores

Remove images annotated with certain "ignore" category labels. This is usually used for removing rubbish images that are pointed out by annotators to ignore frame.

python3 -m cocojson.run.ignore_prune -h

Remove Empty

Remove empty/negative images from COCO JSON, aka images without associated annotations.

python3 -m cocojson.run.remove_empty -h

Sample

Samples k images from a dataset

python3 -m cocojson.run.sample -h

Sample by Category

Samples images from each category for given sample number(s).

python3 -m cocojson.run.sample_by_class -h

Split

Split up a COCO JSON file by images into N sets defined by ratio of total images

python3 -m cocojson.run.split -h

Split by Image Meta-Information

Split up a COCO JSON file by images' meta-information/attributes

python3 -m cocojson.run.split_by_meta -h

Visualise

Visualise annotations onto images. Best used for sanity check.

python3 -m cocojson.run.viz -h

Converters

CVAT Video XML to COCO JSON

Convert CVAT Video XML to COCO JSON whilst preserving track information.

python3 -m cocojson.run.cvatvid2coco -h

CVAT Image XML to COCO JSON

TODO

CrowdHuman odgt to COCO JSON

Converts CrowdHuman's odgt annotation format to COCO JSON format.

python3 -m cocojson.run.crowdhuman2coco -h

Custom Object Detection Logging format to COCO JSON

Converts Custom Object Detection Logging format to COCO JSON format.

python3 -m cocojson.run.log2coco -h

COCO to Darknet