Tips and Traps in Deep Learning Data Handling

While deep learning research naturally focuses on tweaking algorithms and can draw on existing datasets, obtaining enough suitable image data is already a challenge for industrial applications. In this presentation, Gärtner begins by discussing the critical role of robust data pipelines in industrial settings and shows that data-centric approaches often yield better results than model-centric approaches. From there, he explores the complexities of managing datasets for deep learning. Gärtner touches on the different forms of "garbage" that exist in deep-learning-based computer vision and provides practical guidance and insights to optimize the performance of deep learning applications in industrial contexts.

About the presenter
Jan Gärtner has been working at MVTec Software GmbH since 2020, initially as an application engineer. Since early 2023, he has been serving as the product manager for HALCON. He studied electrical engineering and information technology at the Technical University of Munich.