Improved Semi-Supervised Attention GAN for Semantic Segmentation

Semantic segmentation is one of the cornerstone problems in computer vision that involves assigning each image pixel to a specific semantic class. Traditional supervised learning approaches are heavily dependent on labeled data, which is often costly and time-consuming to obtain. Semi-supervised learning approaches, on the other hand, offer a promising path to improve segmentation accuracy by combining labeled and unlabeled data. This work uses an attention-driven adversarial training strategy-based generative adversarial network (GAN) to create realistic semantic segmentation maps for unlabeled data while enhancing segmentation accuracy for labeled data. Additionally, it introduces a patch-wise discriminator to extract rich contextual information. Extensive ablation studies on two widely adopted datasets, Cityscapes and CamVid, demonstrate the effectiveness of the proposed model, achieving state-of-the-art performances. It contributes to the advancement of semi-supervised learning in semantic segmentation, providing a practical solution for improving segmentation accuracy while reducing the reliance on labeled data.