@article{Mueller_TPAMI25,
TITLE = {Examining the Impact of Optical Aberrations to Image Classification and Object Detection Models},
AUTHOR = {M{\"u}ller, Patrick and Braun, Alexander and Keuper, Margret},
LANGUAGE = {eng},
ISSN = {0162-8828},
DOI = {10.1109/TPAMI.2025.3622234},
PUBLISHER = {IEEE},
ADDRESS = {Philadelphia, PA},
YEAR = {2025},
MARGINALMARK = {$\bullet$},
DATE = {2025},
ABSTRACT = {Deep neural networks (DNNs) have proven to be successful in various computer<br>vision applications such that models even infer in safety-critical situations.<br>Therefore, vision models have to behave in a robust way to disturbances such as<br>noise or blur. While seminal benchmarks exist to evaluate model robustness to<br>diverse corruptions, blur is often approximated in an overly simplistic way to<br>model defocus, while ignoring the different blur kernel shapes that result from<br>optical systems. To study model robustness against realistic optical blur<br>effects, this paper proposes two datasets of blur corruptions, which we denote<br>OpticsBench and LensCorruptions. OpticsBench examines primary aberrations such<br>as coma, defocus, and astigmatism, i.e. aberrations that can be represented by<br>varying a single parameter of Zernike polynomials. To go beyond the principled<br>but synthetic setting of primary aberrations, LensCorruptions samples linear<br>combinations in the vector space spanned by Zernike polynomials, corresponding<br>to 100 real lenses. Evaluations for image classification and object detection<br>on ImageNet and MSCOCO show that for a variety of different pre-trained models,<br>the performance on OpticsBench and LensCorruptions varies significantly,<br>indicating the need to consider realistic image corruptions to evaluate a<br>model's robustness against blur.<br>},
JOURNAL = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
VOLUME = {48},
NUMBER = {3},
PAGES = {2139--2153},
}