Doctoral Defense - Minchul Kim

About the Event

The Department of Computer Science & Engineering

Michigan State University

Ph.D. Dissertation Defense

April 21st, 2025 at 4:00pm EST

EB 3105 & https://msu.zoom.us/j/98501814258

Passcode: Upon Request from Vincent Mattison or Advisor

From Pixels to Identity: Visual Recognition and Biometric Applications

By: Minchul Kim

Advisor: Dr. Xiaoming Liu

This thesis explores the evolving field of visual recognition, with a particular emphasis on biometric applications such as face and body identification. It addresses core challenges in the domain, including image degradation, dataset limitations, misalignment, and multimodal recognition. Through a combination of novel algorithmic approaches and comprehensive datasets, the research advances the state of the art in biometric recognition. Key contributions include AdaFace, which employs an adaptive loss function to improve performance on low-quality images, and CAFace, an attention-based feature fusion framework designed for video-based recognition. To tackle misalignment issues, KPRPE introduces keypoint-relative positional encoding. Additionally, SapiensID offers a unified framework for joint face and body recognition, supported by the large-scale multimodal dataset WebBody4M. By enhancing model robustness, data efficiency, and recognition accuracy across diverse real-world conditions, this thesis contributes toward reliable and scalable biometric systems. The work demonstrates how visual recognition technologies can transition effectively from raw pixel data to high-level identity inference in practical applications.