ECCV 2022 - Wednesday

Digital cameras are fundamental to a range of intelligent systems that recognize relevant events and assist us in our activities, with multiple applications in healthcare and beyond. However, their ever-improving ability to imitate the human vision system and produce the highest- quality images has raised concerns about privacy and security . Previous works have focused on software-level processing of videos after they have been filmed to resolve these concerns, but this can leave original videos unprotected. Carlos wants to address this problem within the camera hardware itself. He proposes a new parameterized lens design that adds distortions optimized to preserve privacy while retaining features that allow computers to perform human action recognition tasks. “ The main idea is to design the lens jointly with a human action recognition network and an adversarial network that performs attacks PrivHAR: Recognizing Human Actions from Privacy-Preserving Lens 4 Oral Presentation Carlos Hinojosa just completed his PhD at the Industrial University of Santander in Colombia under the supervision of Henry Arguello. In his final year, he had an internship at the Stanford Vision and Learning Lab, where he worked with Juan Carlos Niebles, Fei-Fei Li, and Ehsan Adeli on a project about preserving privacy to perform computer vision tasks in human action recognition. He speaks to us about that work ahead of his oral presentation this afternoon.