Stony Brook University SUNY Korea

Title: From CCTV to Insights: The Art and Science of Video Analytics

Abstract

In online shopping, tracking and analysis of user statistics is standard practice. Collecting such user behavior statistics in the real world is far more challenging. Video analytics attempts to achieve exactly that by using machine learning algorithms to extract useful information from video data commonly captured through CCTV. Deeping Source Inc., a Seoul-based startup provides video analytics solutions for the retail and theme park industry while protecting individuals' privacy information. This talk peeks behind the curtain of Deeping Source technology into the research efforts necessary for state-of-the-art video analytics and anonymization software. One core technology of Deeping Source is multiview pedestrian detection. Hence the first part of this work will focus on “Booster-SHOT”, an end-to-end convolutional approach to multiview pedestrian detection incorporating a Homography Attention Module as well as view-coherent augmentation or stacked homography transformations. Further, we will investigate how noisy adversarial representation learning can be leveraged for effective and efficient image obfuscation.

Bio

In his current position as the manager of Deeping Sources (Seoul) research team, Benz is leading a group of talented individuals committed to innovating solutions in image anonymization and AI video analytics. Their focus lies at the intersection of multi-camera object tracking, and image anonymization, all with a clear emphasis on practical, real-world applications.

Dr. Philipp Benz received his Ph.D. from the Robotics and Computer Vision Lab at KAIST in South Korea, working under the expert mentorship of Professor Kweon In So. During this period, he delved into the realm of deep learning, specializing in techniques that exemplify the robustness and reliability of Deep Neural Networks. Specifically, he researched adversarial examples, deep neural network robustness to image corruptions, deep image hiding, and the fairness of deep neural networks.