Dense Spatial Translation Network

Abstract

Neural networks are widely used in autonomous driving and driver assistance systems tasks. Limited by hardware, these networks are restricted by their capacity and capability. To deal with this limitation, an application dedicated unit which exploits prior knowledge on beneficial steps may reduce the required network complexity. We introduce a neuralnetwork-integrable unit, Dense Spatial Translation Network (DSTN), that compensates for complex intra-class variations in spatial appearance. For example, considering Traffic Sign Recognition (TSR), the design of the same traffic sign in different countries may be different. This efficient unit is explicitly designed for this rectification task and thus replaces the demand to substantially increase the network capacity. It samples input feature maps which are augmented by intra-class variations, and produces output feature maps compensating for these variations. This clearly simplifies the subsequent classification tasks. Also, the DSTN is light-weighted, and is suitable for end-to-end training. It is easily integrated into any existing network structure. We evaluate the performance of the unit based on TSR and number recognition. Results show significant improvement after integrating this unit into a neural network.