Ruhrgebiet AI & Data Science Meetup Essen
Talk: Symmetries in Visual Recognition
This talk discusses visual recognition, i.e. detection and classification of objects with known geometry in RGBD images. The geometric information inherent in the depth channel is key to a number of applications in industrial robotics, e.g., loading a machine with a raw part picked by vision, where the precise relative position of part, machine, and hence gripper must be known. The knowledge of depth also enables us to better deal with symmetries which are a nuisance to visual recognition in general. Symmetries are generated by imaging one and the same object class from multiple vantage points or under different lighting conditions. After providing a formal definition of all symmetries arising in the image formation process, we show how their consideration in
well-known feature/image descriptors such as SIFT and HOG has pushed the envelope of visual recognition before the advent of (convolutional) neural networks (CNN). We briefly touch upon the question of whether symmstries can be learned by data-driven algorithms such as CNNs
before discussing how different steps of the visual recognition pipeline are implemented within a scalable state-of-the-art microservice architecture.
About the Speaker
Dr. Jonathan Balzer obtained a Master’s Degree in Mechanical Engineering with strong focus on factory automation and robotics from the Karlsruhe Institute of Technology in 2004. During his PhD studies at the department of Computer Science of the same school, he was also enrolled in a
Master’s Program in Applied Mathematics, completing both studies successfully in 2008. From 2008 to 2011, he worked with Prof. Helmut Pottmann in the field of Computational Geometry and Computer Graphics, first at the Technical University of Vienna, later at King Abdullah University of Science and Technology (KAUST) in Thuwal, Saudi Arabia. His final years as a postdoctoral researcher, he spent at the University of California Los Angeles (UCLA) in Prof. Stefano Soatto’s Machine Vision Lab. His work has been published in prestigious conference and journal venues in Computer Vision and Computational Geometry. Jonathan is the author of several open-source software libraries, and has taught college-level Math and Computer Science classes. He returned to Germany in 2015 to found Vathos where he serves as CTO ever since. He is an avid mountaineer and multi-instrumental musician.