Abstract

This talk examines how computer vision methods evolve as data availability. We begin before deep learning, showing how an optical music recognition (OMR) system can be built with classical pipelines when labeled data are scarce. We then move into the deep learning era with context-aware padding prediction for scene parsing in autonomous driving, using data to learn boundary context rather than relying on fixed padding rules. Finally, we discuss applying foundation models to wafer skeleton identification in semiconductor packaging, where events are rare and traceability is strict. Across these domains, we present a practical playbook: match your method to your data regime—add rules when labels are limited, scale and regularize when data are abundant but long-tailed, and leverage pretrained priors when supervision is costly.

Bio.

Yu-Hui Huang is an Assistant Professor with the Department of Electrical Engineering, Yuan Ze University, Taiwan. She received the Ph.D. degree in Electrical Engineering from KU Leuven, Belgium, and the M.Sc. degree from RWTH Aachen University, Germany. Her research focuses on computer vision and deep learning, particularly in scene understanding, motion analysis and optical music recognition.

All faculty and students are welcome to join.