- Wednesday 15 February 2023
Niels Bohrweg 1
2333 CA Leiden
Despite the significant advancements made by deep models in various visual tasks, their success largely relies on the availability of vast amounts of annotated training data. However, annotating data can be a time-consuming and labor-intensive process, particularly in complex tasks such as object detection. To address this challenge, relaxing the labeling cost has become a critical problem in deep learning. In this talk, I will explore a potential solution to this problem through the view of data augmentation. Specifically, I will introduce "Random Erasing," a simple and effective data augmentation method that generates new visual training samples by randomly erasing regions within images. This method is network-agnostic and can be easily applied to various vision tasks, resulting in consistent performance improvement. This method has been included in PyTorch as a data augmentation function.
About Zhun Zhong
Dr. Zhun Zhong received his Ph.D. degree in 2019 from Xiamen University. He is now an assistant professor at the University of Trento and was a postdoc at the same place. He is committed to designing robust and scalable visual recognition systems for real-world applications. His research interests include data augmentation, domain generalization/adaptation, novel class discovery, and object retrieval. He has published more than 20 peer-reviewed papers in top conferences and journals (over 6,700 citations; h-index 20 (GS)), in which two of his first-authored papers have received over 1, 000 citations each. He was an area chair or a senior program committee in several top conferences, e.g., ACM MM, AAAI, and IJCAI. He received the Outstanding Reviewer Award at CVPR 2020 and NeurIPS 2021. He was selected as the AI 2000 Most Inﬂuential Scholar Honorable Mention in AAAI/IJCAI.