ÐÓ°ÉÖ±²¥

Dr Xin Yu is a senior lecturer at the ÐÓ°ÉÖ±²¥ of Queensland. He is an ÐÓ°ÉÖ±²¥n Research Council (ARC) Discovery Early Career Researcher Award 2023-2025 (DECRA) recipient and an awardee of the prestigious Google Research Scholar Program in 2021. He received his PhD degrees from both the ÐÓ°ÉÖ±²¥n National ÐÓ°ÉÖ±²¥ and Tsinghua ÐÓ°ÉÖ±²¥. His research interests involve various Computer Vision and Machine Learning topics, especially in visual data quality enhancement and recovery, human movement understanding, medical imaging analysis, and multimodal data understanding.

In this talk, I will explore the transformative impact of data-centric approaches in computer vision, shifting the focus from model-centric optimization to leveraging the full potential of data. Traditional computer vision pipelines often emphasize algorithmic advancements, but real-world applications demand a deeper understanding of how data quality, and distribution influence model performance. A data-centric perspective enables us to pinpoint challenges in newly collected datasets, uncover hidden biases, and design task-specific algorithms that are both robust and generalizable.

I will illustrate this concept through two case studies. The first focuses on medical imaging, specifically retinal vessel imagery, where high-quality domain-specific data is crucial for accurate disease diagnosis and prognosis. I will discuss the challenges of data variability and domain adaptation, along with strategies to enhance dataset reliability and improve model performance.

The second case study delves into ÐÓ°ÉÖ±²¥n Sign Language (Auslan) translation, a complex task requiring multimodal data integration. I will highlight how a data-centric approach—through better dataset curation, representation learning, and cross-modal alignment—can significantly improve sign language recognition and translation accuracy.

Through these examples, this talk will demonstrate how data-centric computer vision can lead to innovation across various real-world applications.