Real-time Semantic Segmentation for Road Scene

Abstract

Semantic segmentation is challenging for diverse scenes. In this paper, we propose a new model for road complex scene. Our model use encoder-decoder structure with auxiliary loss. Based on ResNet bottleneck block, we proposed dilated bottleneck block and tiny block. These block applied in the encoder and decoder. The dilated bottleneck block enlarges the field-of-view and the tiny block maintains the model as small as possible. We train our model end-to-end from scartch, and the image and segmentation map in network is pixel-to-pixel. With the help of auxiliary loss, our model yields 56.9% mean IoU on CamVid dataset, it is smaller and faster than ENet and SegNet.