roujin/GTA5subset

GTA5 Subset for Zero-Shot Domain Adaptive Semantic Segmentation

This repository contains a curated subset of the GTA5 dataset, specifically designed for experiments in the paper Zero Shot Domain Adaptive Semantic Segmentation by Synthetic Data Generation and Progressive Adaptation. This subset includes images and labels necessary for training and evaluating models in zero-shot domain adaptive semantic segmentation scenarios.

The original GTA5 dataset was extracted from https://download.visinf.tu-darmstadt.de/data/from_games/.

Paper

Zero Shot Domain Adaptive Semantic Segmentation by Synthetic Data Generation and Progressive Adaptation Jun Luo, Zijing Zhao, Yang Liu

Abstract: Deep learning-based semantic segmentation models achieve impressive results yet remain limited in handling distribution shifts between training and test data. In this paper, we present SDGPA (Synthetic Data Generation and Progressive Adaptation), a novel method that tackles zero-shot domain adaptive semantic segmentation, in which no target images are available, but only a text description of the target domain's style is provided. To compensate for the lack of target domain training data, we utilize a pretrained off-the-shelf text-to-image diffusion model, which generates training images by transferring source domain images to target style. Directly editing source domain images introduces noise that harms segmentation because the layout of source images cannot be precisely maintained. To address inaccurate layouts in synthetic data, we propose a method that crops the source image, edits small patches individually, and then merges them back together, which helps improve spatial precision. Recognizing the large domain gap, SDGPA constructs an augmented intermediate domain, leveraging easier adaptation subtasks to enable more stable model adaptation to the target domain. Additionally, to mitigate the impact of noise in synthetic data, we design a progressive adaptation strategy, ensuring robust learning throughout the training process. Extensive experiments demonstrate that our method achieves state-of-the-art performance in zero-shot semantic segmentation. The code is available at this https URL

Code

The official implementation for the paper can be found on GitHub: https://github.com/roujin/SDGPA

Dataset Structure and Usage

To use this GTA5 subset with the official SDGPA codebase, you should organize your datasets under a common data root (e.g., <data_root>). This dataset specifically forms the GTA5 folder within that structure.

An example of the required folder organization:

<data_root>
└─ ACDC
   └─ gt
   └─ rgb_anon
└─ cityscapes
   └─ gtFine
   └─ leftImg8bit
└─ GTA5
   └─ images
   └─ labels

For more details on installation and running experiments, please refer to the GitHub repository.

License

This dataset is licensed under the MIT License. More details can be found at: https://bitbucket.org/visinf/projects-2016-playing-for-data/src/master/LICENSE.md

Citation

If you use this dataset in your research, please cite the associated paper:

@misc{luo2025sdgpa,
      title={Zero Shot Domain Adaptive Semantic Segmentation by Synthetic Data Generation and Progressive Adaptation}, 
      author={Jun Luo and Zijing Zhao and Yang Liu},
      year={2025},
      eprint={2508.03300},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.03300}, 
}