Dynamic Split Computing of PoseNet Inference for Fitness Applications in Home IoT-Edge Platform

• Published in: 2022 14th International Conference on COMmunication Systems & NETworkS (COMSNETS) pp. 430 - 432
• Main Authors: Karjee, Jyotirmoy, Anand, Kartik, S, Praveen Naik, Dabbiru, Ramesh Babu Venkat, Byadgi, Chandrashekhar S, N, Srinidhi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04-01-2022
• Subjects: Bandwidth; Computational modeling; Economic indicators; Heuristic algorithms; Inference algorithms; Partitioning algorithms; Performance evaluation
• ISSN: 2155-2509
• DOI: 10.1109/COMSNETS53615.2022.9668605

In the next generation wireless networks, implementations of advanced computation tasks such as running Deep Neural Network (DNN) models in Internet of Things (IoT) devices is a challenging task due to their limited computation & processing capabilities. To address this issue, we propose an Extended Dynamic Split Computation (E-DSC) algorithm which finds an optimal partition point of DNN inference layer based on the available network bandwidth; where one inference sub-model is computed at IoT device and the other inference sub-model is computed by a home edge. To validate the E-DSC algorithm, we consider a fitness application as a use-case. The application captures live video of a person performing an arm exercise using a camera attached to a Raspberry-Pi (RPi) considered as an IoT device. The PoseNet DNN model is used to estimate the pose of the person using split-inference of the DNN are partitioned among RPi and Samsung Galaxy S20 device (considered as a home edge). The S20 device sends the final inference result back to the RPi which then checks if the exercise is done correctly and displays the exercise count on a television/monitor. We conduct extensive experiments to compare the performance of PoseNet inference execution on RPi device, Galaxy S20 device and performing dynamic split using E-DSC algorithm over Wi-Fi networks in a real time deployment scenario.