Robust Visual Perception in Adverse Conditions

Vision models operating in the real world often face adverse conditions such as severe weather or low light. Their robustness against input degradation is thus crucial for reliable deployment. This talk introduces a series of studies aimed at ensuring the robustness of vision models against adverse conditions, focusing on three points: (1) learning representations invariant to input degradations, (2) input-aware dynamic rectification of vision models for on-the-fly robustification, and (3) training-data collection using specialized camera systems or degradation-synthesis pipelines. Finally, I discuss key considerations and future directions for robust vision models.

Speaker

Suha Kwak is an Associate Professor in the Graduate School of Artificial Intelligence at POSTECH. He received his Ph.D. from POSTECH and was a Postdoctoral Fellow at École Normale Supérieure / Inria, Paris. He is a recipient of the Kakao Faculty Fellowship (2018) and the KCCV Sang-Uk Lee Prize (2023); one of his papers was a Best Paper Finalist at CVPR 2022, and he was named an Outstanding Area Chair at ECCV 2024. He served as Associate Editor of IJCV and as (lead) Area Chair for CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML and AAAI.