SSIGAN: Semantic Segmentation via Improved Generative Adversarial Network

Abstract

Nowadays, although conditional convolutional neural networks have applied to semantic segmentation, their loss function needs to be carefully designed. We propose an improved generative adversarial network including a generator network and a discriminator network for semantic segmentation. In some blocks, we substitute 3×1 and 1×3 factorized convolution for 3×3 convolution to make full use of transverse and longitudinal information. We concat the original image with the output of the generator as the input of the discriminator network to improve the discriminant ability. As a result, our model achieves 69.6% mean intersection over union (mIoU) on the Cityscapes test set. Our experiments exhibit that adversarial training approach leads to improved accuracy on the Cityscapes, Camvid, Kitti and Gatech dataset in road scene.