openimages2coco

Convert Open Images annotations into MS Coco format to make it a drop in replacement.

Functionality

• convert_annotations.py will load the original .csv annotation files from Open Images, convert the annotations into the list/dict based format of MS Coco annotations and store them as a .json file in the same folder.

• convert_predictions.py loads a .json file with predictions in the coco format and save them as .csv in the OpenImages prediction fromat at the same location.

Installation

Download the CocoAPI from https://github.com/cocodataset/cocoapi
Install Coco API:

cd PATH_TO_COCOAPI/PythonAPI
make install

Download Open Images from https://storage.googleapis.com/openimages/web/download.html
-> Store the images in three folders called: train, val and test
-> Store the annotations for all three splits in a separate folder called: annotations

Converting Annotations

Run conversion of bounding box annotations:

python3 convert_annotations.py -p PATH_TO_OPENIMAGES --task bbox

The convert instance masks to the Coco panoptic format. The masks have to be placed in annotations/SPLIT_masks/

python3 convert_annotations.py -p PATH_TO_OPENIMAGES --task panoptic

Currently adding the instance masks to the annotations as done for coco is not supported becasue the resulting json file would be extremely large.

Converting Predictions

Run conversion of bounding box predictions:

python3 convert_predictions.py -p PATH_TO_PREDICTIONS --subset validation

The subset is necessary to get the correct image sizes. Alternatively they can be inferred directly from the images:

python3 convert_predictions.py -p PATH_TO_PREDICTIONS --image_dir PATH_TO_IMAGES

Currently only bounding box predictions are supported.

Dataset Versions

The toolkit supports multiple versions of the dataset including v4, v5, v6 and challenge_2019. For example the bbox annotations of challenge_2019 can be converted like:

python3 convert_annotations.py -p PATH_TO_OPENIMAGES --version challenge_2019 --task bbox

For panoptica nnotations the masks have to be placed in annotations/challenge_2019_$SPLIT_masks/ before running:

python3 convert_annotations.py -p PATH_TO_OPENIMAGES --version challenge_2019 --task panoptic

Note, that different annotation files have to be downloaded to annotations for this purpose. The files for the challenge_2019 set can be found here: https://storage.googleapis.com/openimages/web/challenge2019_downloads.html

Using Converted Annotations

The generated annotations can be loaded and used with the standard MS Coco tools:

from pycocotools.coco import COCO

# Example for the validation set
openimages = COCO('PATH_TO_OPENIMAGES/annotations/openimages_v6_val_bbox.json')