April 1, 2020

News: Manuscript is avaiable in Arxiv.

• The code works with **PyTorch 1.0.0 or higher.

This repository provides bounding box annotations of the ObjectNet dataset by Barbu et al. NIPS 2019.

About the ObjectNet dataset: Taken from the ObjectNet repository README.md, verbatim: "ObjectNet is an object recognition dataset developed at MIT in CSAIL & BCS as well as CBMM (Center for Brains, Minds, and Machines) and at the MIT-IBM Watson AI Lab. It intentionally adds variation to the background, viewpoint, and rotation of objects during the image capturing process while minimizing correlations between these properties and object class. Removing these biases challenges current object detectors. We hope that this will lead to new detectors and new ideas in machine learning as well a more rigorous and scientific approaches to datasets everywhere.

ObjectNet has no training set! The ObjectNet license, attached, which you agree to in order to use this work, EXPLICITLY FORBIDS UPDATING THE PARAMETERS OF ANY MODEL BASED ON THESE IMAGES. Our goal is for ObjectNet and datasets like it to produce accuracies that are predictive of what any user, including you, will see in the real world when deploying object detectors."

** passsword for the .zip file: objectnet1234

If you use ObjectNet Annotations or found the ideas in the paper useful, please cite:

ObjectNet Dataset: Reanalysis and Correction Ali Borji

Here is the .bib file:

@misc{borji2020objectnet,
    title={ObjectNet Dataset: Reanalysis and Correction},
    author={Ali Borji},
    year={2020},
    eprint={2004.02042},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}


@misc{borji2021Contemplating,
    title={Contemplating real-world object classification},
    author={Ali Borji},
    year={2021},
    jounral={ICLR},
    primaryClass={cs.CV}
}

Please write to [email protected] for further questiones, comments, etc.