Education Class 9

Session: EC9

Educator: Soonhoi Ha

Institute: Seoul National University

Time & Date: September 15, 2023, 10:00 – 12:00 CET

Title: Design Methodology for Low Power Computer Vision Systems

Abstract: In the development of an embedded computer vision system, there are many issues to consider, such as which hardware platform and algorithm to use, how to optimize the software with resource constraints and how to optimize multiple design objectives, and so on. This lecture presents a systematic design methodology that could be applied to the design of embedded computer vision systems. Based on the proposed methodology, we could win the first prize in LPIRC (Low-Power Image Recognition Challenge) 2017 and in track 2 of LPIRC 2018 using a NVIDIA Jetson TX2 board. Recently new hardware platforms have been developed that contain CNN (Convolutional Neural Network) hardware accelerators as well as GPU (Graphics Processing Units), among which NVIDIA Jetson AGX Xavier is a representative example. Since it is a heterogeneous system that contains multiple hardware accelerators, exploiting the computing power of those accelerators maximally becomes an important issue to consider in the proposed design methodology. I will present a novel technique to support multiple deep learning applications as well as other dataflow applications on such a heterogeneous multi-processor platform.