説明
The 50th Seminar
Date and Time: January 24th 5:00 pm – 6:00 pm(JST)
Venue: Zoom webinar
Language: English
Speaker: Jenny Schmalfuss (University of Stutgart)
Title: Challenges in Adversarial Attacks for Motion Estimation
Short Abstract:
When you see an approaching train, you directly register that it is moving towards you. Even if the scene changes a little, because the glasses you are wearing are dusty or because snowflakes are blown into your view, this will not lead you to believe that the train changed its direction or disappeared. What seems obvious to us as humans is challenging for algorithms that estimate 2D motion from videos, also known as the optical flow problem. While current methods for optical flow estimation achieve an impressive quality for their predictions, they are also very susceptible to adversarial attacks that inconspicuously change the video frames to cause dramatically wrong motion predictions. In this talk, we investigate the robustness of optical flow methods with the help of adversarial attacks, and discuss the differences in motion robustness among attacks with Lp perturbations (Schmalfuss et al. ECCV 2022) and attacks with non-Lp perturbations in the form of photorealistic adversarial snow (Schmalfuss et al. ECCV-AROW 2022).
Bio:
Jenny Schmalfuss is a PhD student at the University of Stuttgart, and a fellow of the International Max Planck Research School for Intelligent Systems (IMPRS-IS). Her research interests are at the intersection of computer vision and machine learning. In her PhD she investigates new ways to quantify and improve the robustness of motion estimation methods with the help of adversarial attacks. Her goal is to enable robust motion estimation by understanding what influences the robustness of current methods.
