On the acceleration of the convolutional neural network architecture based on ResNet in the task of road scene objects recognition

Recent approaches to road scene objects detection based on convolutional neural networks have reached an acceptable level to be used in autonomous vehicle control and ADAS systems. However, the best modern network architectures are rather heavy and cannot be integrated in real-time systems. Thus the most actual problem is to accelerate networks and to find the optimal balance between their quality and performance. This paper proposes a method to facilitate the architecture of the Deformable Convolutional Network based on ResNet backbone that provides a threefold increase in the inference performance. At the same time, the quality of detection of road scene objects is reduced not so significantly. In addition, the paper compares the quality of the network of this architecture trained on different open datasets – BDD and MS-COCO.