This repository points to the official implementation for our paper:
ImageNet-LC: Assessing Robustness under Localized Corruptions
ICPR 2026
ImageNet-LC is an object-centric robustness benchmark for ImageNet. Instead of applying corruptions globally to the full image, ImageNet-LC localizes the foreground object regions and applies corruptions only inside those regions of interest. This enables evaluation of model robustness under more targeted, localized visual degradations.
The benchmark includes localized corruptions such as lens flare, illumination variation, dust and scratches, fingerprint smudges, focus shift, occlusion, and camouflage across multiple severity levels.
The main code release is available here:
https://github.com/muskanny/ImageNet-LC
Please use the main repository for installation instructions, pipeline usage, updates, and issue tracking.
The ImageNet-LC dataset is available on Kaggle:
https://www.kaggle.com/datasets/sanchitgupta10/imagenet-pc
If you use ImageNet-LC in your work, please cite the paper:
@inproceedings{imagenetlc2026,
title = {ImageNet-LC: Assessing Robustness under Localized Corruptions},
booktitle = {ICPR},
year = {2026}
}