FAST: Fast and Accurate Adaptation in Live Video Analytics Using Intermediate Features.

Live video analytics need to adapt the optimal configuration to achieve the requirements for various videos of environments. Prior works for adaptation suffer from the accuracy drop due to the large overhead and incorrect estimation. We present FAST, a Feature-based Adaptation SysTem, for immediate and accurate adaptation in live video analytics. To overcome limitations of prior works, we leverage the intermediate feature of the task. Intermediate features enable fine-grained adaptation and ambiguity distinction, which leads to reducing the frame size to be transmitted and increasing the accuracy. We design end-to-end systems for two representative tasks (object detection and heartrate measurement) to verify the benefits of the intermediate features. Evaluations show FAST achieves up to 6.95 x lower average frame size and 2.69 x higher accuracy compared to state-of-the-art adaptation.