GStreamer Video Analytics: Optimizing inference across HW targets. Neelay Shah, Neena Maldikar, Mikhail Nikolsky & Ilya Belyakov, Intel

Loading
Loading Click here to add:
Add to notification list

Intel has recently open sourced a set of video analytics plugins based on the OpenVINO inference engine. These plugins provide an easy way to add neural net based analytics (object detection, classification, identification, etc.) to GStreamer pipelines in a way that is optimized across different HW types (CPU, GPU, VPU, FPGA). In this talk we will describe how the plugins are designed and how to use and customize them in applications. We are excited to demonstrate these new capabilities as well as solicit feedback from the community on the future roadmap. Neelay Shah is a software architect at Intel developing video analytics applications using GStreamer for use cases in smart cities, retail and broadcasting. Graduating from the University of Illinois at Urbana Champaign in 2006 with a master’s degree in computer science, he has worked at Intel for over 10 years on various projects including UPnP, context sensing and most recently visual computing. Neena Maldikar is the product owner for GStreamer Video Analytics. She has a master’s degree in Computer Science from Portland State University. She has been working at Intel for the past 6 years. Before joining the Video Analytics team, she worked on bringing contextual awareness to PC platforms through new libraries, applications and the integration of new sensors. Mikhail Nikolskii is the lead architect and developer of GStreamer Video Analytics. He has worked at Intel for over 15 years on various Intel software products. Mikhail has a master’s degree in Computer Science from Moscow State University Ilya Belyakov is the engineering manager for GStreamer Video Analytics. He graduated from the State University of Nizhni Novgorod with a degree in computer science. He has worked at Intel for the past 7 years on various technologies including Intel(R) RealSense(TM) and advanced labeling tools to train autonomous driving algorithms.