Fast and Accurate RMNet: A New Neural Network for Embedded Vision

Usually, the top places in deep learning challenges are won by huge neural networks that require massive amounts of data and computation, making them impractical for use in real-time edge applications like security and autonomous driving.

In this talk we will describe a new neural network architecture, RMNet, designed to achieve a balance of accuracy and performance for embedded vision applications. RMNet merges the best practices of network architectures like MobileNets and ResNets. To demonstrate the effectiveness of this new network, we present an evaluation of RMNet on a person reidentification task. In terms of accuracy, the proposed approach takes third place on the Market-1501 challenge, while offering much better inference speed. RMnet can be used for many tasks such as face recognition, pedestrian detection, vehicle detection and bicycle detection.