S. Zehtabian, S. Khodadadeh, L. Bölöni, and D. Turgut. Improving AR/VR experiences with deep learning. In Proc. of 1st Annual Nelms Workshop on Women in IoT (WiT-2020), October 2020.
In the near future, augmented/virtual reality (AR/VR) characters might be used for tasks in the home that are currently performed through cell phones or laptop computers - ranging from checking the weather or news to performing banking, visiting a doctor, or going to school. Instead of a keyboard or touch screen interface, the user will interact with a virtual or real person, visualized life-size, with high quality through large screens or AR/VR devices. User satisfaction for such applications depends on delivering high-quality content with minimum latency. In this poster we describe a technique where we predict the user's future requests, use the prediction to prefetch the data from the network, cache it on a local device and show it to the user at the right time with minimum latency and maximum quality. We describe a deep learning technique to predict the AR/VR experiences that the users are most likely to access at a specific time of the day and develop several different caching techniques. We rely on real-world smart home datasets, augmented with synthetic data created to match the essential attributes of the real-world data. We evaluate the proposed prediction methods and calculate the user's experience scores in terms of caching costs and user satisfaction. Finally, we compare our results with other baselines such as random caching, caching everything, and oracle. We found that our predictive approaches outperform the baselines, the difference being especially significant for the high-quality format deliveries.
@inproceedings{Zehtabian-2020-WiT,
author = "S. Zehtabian and S. Khodadadeh and L. B{\"o}l{\"o}ni and D. Turgut",
title = "Improving AR/VR experiences with deep learning",
booktitle = "Proc. of 1st Annual Nelms Workshop on Women in IoT (WiT-2020)",
year = "2020",
month = "October",
abstract = {
In the near future, augmented/virtual reality (AR/VR) characters might be used for tasks in the home that are currently performed through cell phones or laptop computers - ranging from checking the weather or news to performing banking, visiting a doctor, or going to school. Instead of a keyboard or touch screen interface, the user will interact with a virtual or real person, visualized life-size, with high quality through large screens or AR/VR devices. User satisfaction for such applications depends on delivering high-quality content with minimum latency. In this poster we describe a technique where we predict the user's future requests, use the prediction to prefetch the data from the network, cache it on a local device and show it to the user at the right time with minimum latency and maximum quality.
We describe a deep learning technique to predict the AR/VR experiences that the users are most likely to access at a specific time of the day and develop several different caching techniques. We rely on real-world smart home datasets, augmented with synthetic data created to match the essential attributes of the real-world data. We evaluate the proposed prediction methods and calculate the user's experience scores in terms of caching costs and user satisfaction. Finally, we compare our results with other baselines such as random caching, caching everything, and oracle. We found that our predictive approaches outperform the baselines, the difference being especially significant for the high-quality format deliveries.
},
}
Generated by bib2html.pl (written by Patrick Riley, Lotzi Boloni ) on Sun Mar 03, 2024 18:41:16