This paper examines the influence of detect-and-remove defenses, ILP and LGS, on optical flow methods. Adversarial patches, placed in arbitrary scene locations, can undermine the reliability of optical flow predictions. The study reveals that these defenses not only lower the optical flow quality on benign scenes but also reduce robustness under patch attacks for all tested optical flow methods except FlowNetC. The authors conclude that existing defenses could give a false sense of security as they fail to deliver the promised adversarial robustness for optical flow.
Publication date: 27 Oct 2023
Project Page: https://github.com/cv-stuttgart/DetectionDefenses
Paper: https://arxiv.org/pdf/2310.17403
