A Perturbation-Constrained Adversarial Attack for Evaluating the Robustness of Optical Flow

Recent optical flow methods are almost exclusively judged in terms of accuracy, while their robustness is often neglected. Although adversarial attacks offer a useful tool to perform such an analysis, current on focus real-world attacking scenarios rather than worst case assessment. Hence, this work, we propose novel attack—the Perturbation-Constrained Flow Attack (PCFA)—that emphasizes destructivity over applicability as attack. PCFA global attack that optimizes perturbations shift the predicted towards specified target flow, keeping $$L_2$$ norm perturbation below chosen bound. Our experiments demonstrate PCFA’s white- and black-box settings, show it finds stronger samples previous attacks. Based these strong samples, provide first joint ranking considering both prediction quality robustness, which reveals state-of-the-art be particularly vulnerable. Code available at https://github.com/cv-stuttgart/PCFA .