The Masked MS-COCO dataset is collected from the MS-COCO dataset that is licensed under a Creative Common Attribution 4.0 License. It is created for robust image classification with the hypothesis that background elimination could make the adversarial attack harder by reducing the attack surface, and thus improve the robustness of neural network models.
This dataset has two versions:
- The 6-class version consists of 19,200 32x32 color images from 6 classes ( train, bird, cat, dog, toilet, clock).There are 18,000 training images with 3,000 images per class, and 1,200 test images with 200 images per class.
- The 10-class version consists of 25,000 32x32 color images from 10 classes ( person, airplane, train, bird, dog, elephant, zebra, giraffe, toilet, clock). There are 20,000 training images with 2,000 images per class, and 5,000 test images with 500 images per class.
- We use the Matterport implementation of Mask-RCNN to extract objects and their masks from the MS-COCO dataset. We then resize cropped objects and masks to 32x32, and pad zeros if necessary.
If you use this dataset, please cite this repository (bibtex below).
-
Dataset can be downloaded from Google Drive.
-
Both datasets contain an image directory, a mask directory, and a csv file listing the train and test partitions.
-
Download size:
The 6-class version: 43.2 MB
The 10-class version: 66.3 MB
Use this bibtex to cite this repository:
@misc{cong2019masked,
title={Masked {MS-COCO} for robust image classification},
author={Cong, Tianji and Prakash, Atul},
year={2019},
publisher={Github},
journal={GitHub repository},
howpublished={\url{https://github.com/superctj/Masked_MSCOCO}},
}