Miroslaw Bober – “CNN Architectures for Object Recognition and Visual Search”
“CNN Architectures for Object Recognition and Visual Search”
Abstract: Visual search and specific object recognition are long-standing challenges in Computer Vision and Artificial Intelligence. Recently the area has been significantly advanced by deep learning. This talk will focus on the latest CNN architectures for robust recognition and retrieval. We will start with a brief introduction to core concepts and techniques, including interest points, local and global image descriptors, geometric verification and efficient matching for large-scale visual search systems. We will then move to CNN architectures and introduce the REMAP global image descriptor, which won the Google Landmark Retrieval Challenge on Kaggle in 2018. I will also present the core ideas behind the ACTNET: our latest CNN network with a novel activation layer that defines the state-of-the-art for recognition. Finally, we will briefly look at some applications, including catching criminals, augmenting paper (a-book) and precise self-localisation via recognition.
Bio: Miroslaw Bober is a Professor of Video Processing at the University of Surrey, U.K. where he leads the Media AI team. In 2011 he cofounded Visual Atoms Ltd, a company specializing in visual analysis and search technologies. Between 1997 and 2011 he headed Mitsubishi Electric Corporate R&D Center Europe (MERCE-UK). He received BSc degree from AGH University of Science and Technology, and MSc and PhD degrees from University of Surrey. His research interests include computer vision, machine learning and AI, with a focus on analysis and understanding of visual and multimodal data, and large-scale image and video search. Mirek led the development of ISO MPEG standards for over 20 years, chairing the media analysis groups: MPEG-7, CDVS and CVDA. He is an inventor of over 80 patents, many deployed in products. His publication record includes over 100 refereed publications, including three books and book chapters, and his visual search technologies won the Google Landmark Retrieval Challenge on Kaggle in 2018.