Semantic-aware short path adversarial training for cross-domain semantic segmentation.

Recently, many methods have been proposed to deal with the problem of cross-domain semantic segmentation. Most of them choose to conduct domain adversarial training either on the high-level convolutional neural network (CNN) features or on the output segmentation maps. Typically, a relatively small weight is given to the adversarial training loss to avoid the problem of mode collapse. However, one potential weakness of these methods is that low-level CNN layers may receive little gradients for domain adaptation, especially when the network is deep. To address this problem, we propose to conduct an auxiliary adversarial training on the fused multi-level CNN features. Gradients for domain adaptation can thus flow into low-level CNN layers more easily along a shorter path. Experiments are conducted on the dataset of Cityscapes with using the source datasets of GTA5 and SYNTHIA, respectively. Quantitative and qualitative results certify the efficacy of the proposed method. [ABSTRACT FROM AUTHOR]