Research on Deploying the Deeplearning Models with Embedded Devices

摘要

Deep convolution network has done many awesome tasks in computer vision, showing its ability to go beyond the human, also shining on lots of applications. But high precision always come along with high resource occupation, both of compute resource and memory resource. How to deploy models with consuming resource as less as possible and complete the inference time as short as possible, is becoming an important part of deep learning research. Base on a generic hardware Raspberry 3B+, this paper implements three different ways to deploy a detector on it. The methods are kernel optimization, quantization and the use of movidius VPU. After comparing the results and inference time of different methods, we gave our analysis and summaries.