LAUSR

Visual computing, e.g., automatic object detection, in mobile devices attracts more and more attention recently, in which fast models at constrained energy cost is a critical problem. In this paper, we introduce our work on designing models based on deep learning for 200 classes object detection in mobile devices, as well as exploring trade-off between accuracy and energy cost. In particular, we investigate several methods of extracting object proposals and integrate them into the fast-RCNN framework for object detection. Extensive experiments are conducted using the Jetson TK1 SOC platform and the Alienware-15 laptop, including detailed parameters evaluation with respect to accuracy, energy cost and speed. From these experiments, we conclude how to obtain good balance between accuracy and energy cost, which might provide guidance to design effective and efficient object detection models on mobile devices.