A simple preprocessing approach for improving semantic segmentation in unsupervised domain adaptation

Unsupervised Domain Adaptation (UDA) is a powerful strategy for bridging the gap between synthetic (source) data and real-world (target) data, thereby reducing expensive manual annotations. In this work, we propose ProCST, a novel preprocessing framework that translates source images into target-like images while preserving essential semantic content. Unlike conventional image-to-image or adversarial-based approaches, ProCST utilizes a multi-scale architecture and a dedicated combination of losses-including a new cyclic label loss-to maintain class structure and context. By seamlessly integrating ProCST as a pre-processing stage into existing UDA pipelines, we not only reduce the domain gap but also achieve consistent performance gains. For example, our method improves the mean Intersection-over-Union (mIoU) of state-of-the-art UDA techniques by up to 1.1% on standard tasks such as GTA5 → Cityscapes and 2.2% on an industrial waste segmentation challenge, outperforming current best results. These enhancements underscore ProCST's ability to generate target-like images that retain sufficient semantic fidelity for robust model training. Overall, ProCST offers a cost-effective solution to domain adaptation in semantic segmentation, helping advance real-world applications that rely on large-scale annotated data. Our code and data are available at https://github.com/shahaf1313/ProCST .