Efficient Computer Vision for Embedded Systems
This virtual roundtable collects the opinions from experts in efficient CV about the status of technologies and directions for future improvements as well as evaluating progress in competitions that compare different solutions using the same data sets.
Abstract
decade’s impressive improvements of learning-based artificial intelligence (AI). It also remains one of the grand challenges in AI, where significant R&D effort is required to achieve CV’s fullest potential. One of the driving forces assessing progress in CV is competitions that compare different solutions using the same data sets. Most CV competitions focus on accuracy, without the consideration of ef f ic ienc y on h a rdwa re w it h limited resources. As a result, researchers use increasingly deeper neural networks (NNs), running on fast computers (sometimes supercomputers) with one or more GPUs. Since 2015, the IEEE Low-Power Computer Vision Challenge (LPCVC) has compared CV solutions running on battery-powered devices such as mobile phones and miniature autonomous robots. Over the years, 108 teams from around the world have submitted more than 500 solutions for CV problems including object detection, image classification, moving-object tracking, and character recognition. This virtual roundtable collects the opinions from experts in efficient CV about the status of technologies and directions for future improvements. LPCVC was called the Low-Power Image Recognition Challenge (LPIRC) in 2015–2019. It was renamed LPCVC in 2020 when a video track was added. Efficient Computer Vision for Embedded Systems