NRTK: an open source natural robustness toolkit for the evaluation of computer vision models

Brian Hu; Brandon RichardWebster; Paul Tunison; Emily Veenhuis; Bharadwaj Ravichandran; Alexander Lynch; Stephen Crowell; Alessandro Genova; Vicente Bolea; Sebastien Jourdain; Austin Whitesell

doi:10.1117/12.3012769

7 June 2024 NRTK: an open source natural robustness toolkit for the evaluation of computer vision models

Brian Hu, Brandon RichardWebster, Paul Tunison, Emily Veenhuis, Bharadwaj Ravichandran, Alexander Lynch, Stephen Crowell, Alessandro Genova, Vicente Bolea, Sebastien Jourdain, Austin Whitesell

Author Affiliations +

Proceedings Volume 13054, Assurance and Security for AI-enabled Systems; 130540C (2024) https://doi.org/10.1117/12.3012769
Event: SPIE Defense + Commercial Sensing, 2024, National Harbor, Maryland, United States

Abstract

Recent progress in computer vision has been fueled by advances in artificial intelligence (AI) using techniques such as deep learning. However, computer vision models are often sensitive to distribution shifts, such as those naturally occurring in real operational data, which can degrade performance and cause unforeseen failure modes. As such, it is critical to perform comprehensive test and evaluation of these models across a wide range of natural effects in order to assure their downstream use and deployment. However, it is often difficult, impractical or simply impossible to collect and annotate datasets incorporating the expected range of conditions in operational domains. Here, we introduce the open source natural robustness toolkit (NRTK), a platform for evaluating the natural robustness of computer vision models through realistic augmentations of datasets. The NRTK package provides a set of validated scene- and sensor-specific perturbations using physics-based modeling of the entire image formation process, covering pre-sensor, at-sensor, and post-sensor perturbations. We demonstrate its modular and flexible nature through the example use cases of evaluating object detection models on satellite and unmanned aerial vehicle (UAV) imagery, where critical sensor parameters such as focal length and aperture diameter can be varied to assess model performance. We also show how the NRTK package can be paired with visualization tools to support the interactive exploration of datasets and their perturbations via NRTK. Our results suggest that natural robustness is a critical dimension for evaluation towards the verification and validation of computer vision models, ensuring their trusted use and deployment. Our code is publicly available at: https://github.com/Kitware/nrtk.

Conference Presentation

(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Brian Hu, Brandon RichardWebster, Paul Tunison, Emily Veenhuis, Bharadwaj Ravichandran, Alexander Lynch, Stephen Crowell, Alessandro Genova, Vicente Bolea, Sebastien Jourdain, and Austin Whitesell "NRTK: an open source natural robustness toolkit for the evaluation of computer vision models", Proc. SPIE 13054, Assurance and Security for AI-enabled Systems, 130540C (7 June 2024); https://doi.org/10.1117/12.3012769

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available