Fully combined convolutional network with soft cost function for traffic scene parsing

Abstract

Autonomous car has achieved unprecedented improvement in object detection because of the high performance of deep convolutional neural networks, and now researches are devoted to more complex traffic scene parsing. In this paper, we present a novel traffic scene parsing algorithm by learning a fully combined convolutional network (FCCN). Our network improves the upsampling layer of a fully convolutional network, we add five unpooling layers after the final convolution layer, and each unpooling layer is corresponded to a former pooling layer. We then combine each pair of pooling and unpooling layers, add convolution layers after the combined layer. Since we find it is still hard to learn fine details or edge features of target objects, we propose a soft cost function for further improvement. Our cost function adds soft weights on different target objects. The weight of background is set as constantly one, and the weights for target objects are calculated dynamically, which should be larger than two. We evaluate our work on CamVid datasets. The results show that our FCCN achieves a considerable improvement in segmentation performance.