Vijayan K. Asari
Professor, University of Dayton
Talk Title: Automated data annotation methodology for deep-learning based object detection and recognition
VIJAYAN K. ASARI is a Professor of electrical and computer engineering and the Ohio Research Scholars Endowed Chair in wide area surveillance with the University of Dayton. He is also the Director of the University of Dayton Vision Lab, a center of excellence for computational intelligence and machine vision. He received the Ph.D. degree in electrical engineering from the Indian Institute of Technology, Madras, in 1994. Dr. Asari has so far supervised 30 Doctoral dissertations and 45 Masters theses, and published more than 700 research articles, including an edited book on wide area surveillance and 116 peer-reviewed journal articles in the areas of image processing, computer vision, pattern recognition, machine learning, deep learning, and artificial neural networks. He received several awards for teaching, research, advising and technical leadership that include the University of Dayton Vision Award for Excellence in August 2017, the Sigma Xi George B. Noland Award for Outstanding Research in April 2016, and the Outstanding Engineers and Scientists Award for Technical Leadership from The Affiliate Societies Council of Dayton in April 2015. He is a senior member of IEEE and an elected Fellow of SPIE, and a co-organizer of several IEEE and SPIE conferences and workshops.
Recent advances in sensor technologies have made it possible for high-speed and high-resolution data acquisition resulting in a massive amount of data available for processing. Deep-learning based automatic target recognition algorithms demand the requirement of a large amount of precisely annotated data for robust training and performance evaluation. This necessitates the need for an automated methodology for data annotation. We developed a new semi-supervised deep-learning framework for automated data annotation based on a single-shot object detection algorithm. The proposed data labeling model employs a pre-trained object detection algorithm that could predict the locations of the objects in the images. It automatically creates bounding boxes around the object regions and marks their respective object labels. The annotation tool provides accessibility to finetune the object bounding boxes as needed. The object detection method employed in this model is invariant to object sizes, object orientations, lighting conditions and camera poses. The proposed data annotation system has the capability to retrain itself for capturing information of unfamiliar or new objects for further labeling. It can perform real-time labeling of data captured by the sensors in mobile platforms.