Self-Ensembling GAN for Cross-Domain Semantic Segmentation

Deep neural networks (DNNs) have greatly contributed to the performance gains in semantic segmentation. Nevertheless, training DNNs generally requires large amounts of pixel-level labeled data, which is expensive and time-consuming collect practice. To mitigate annotation burden, this paper proposes a self-ensembling generative adversarial network (SE-GAN) exploiting cross-domain data for In SE-GAN, teacher student constitute model generating segmentation maps, together with discriminator, forms GAN. Despite its simplicity, we find SE-GAN can significantly boost enhance stability model, latter common barrier shared by most training-based methods. We theoretically analyze provide an $\mathcal O(1/\sqrt{N})$ generalization bound ($N$ sample size), suggests controlling discriminator's hypothesis complexity generalizability. Accordingly, choose simple as discriminator. Extensive systematic experiments two standard settings demonstrate that proposed method outperforms current state-of-the-art approaches. The source code our available online (https://github.com/YonghaoXu/SE-GAN).