license: apache-2.0
tags:
- geospatial
task_categories:
- image-segmentation
Dataset Card for Geo_dataset
The Geo_dataset is a cross-domain segmentation dataset for remote sensing images, comprising four subsets: ISPRS Potsdam, ISPRS Vaihingen, Loveda Urban, and Loveda Rural. It's designed for conducting domain adaptation and domain generalization experiments in the context of the Earth-Adapter model. The dataset consists of 512x512 image patches and their corresponding segmentation masks.
Dataset Details
- Paper: Earth-Adapter: Bridge the Geospatial Domain Gaps with Mixture of Frequency Adaptation
- Code: https://github.com/VisionXLab/Earth-Adapter
- Dataset Sources: ISPRS Potsdam, ISPRS Vaihingen, Loveda Urban, and Loveda Rural. Available via Baidu Cloud, Hugging Face, and Google Drive links provided in the GitHub repository.
Uses
Direct Use
The Geo_dataset is intended for research on domain adaptation and domain generalization in remote sensing image segmentation. It can be used to train and evaluate image segmentation models, particularly those focused on handling domain shifts in geospatial data.
Out-of-Scope Use
While the dataset can be used for general image segmentation tasks, its primary focus is on geospatial data and the challenges associated with domain adaptation and generalization in this domain. Therefore, its direct applicability to other domains may be limited.
Dataset Structure
The dataset is organized into four subsets, each containing a set of 512x512 pixel images and their corresponding segmentation masks. Further details on the structure are available within the GitHub repository.
(The remaining sections – Dataset Creation, Bias, Risks, and Limitations, Citation, Glossary, More Information, Dataset Card Authors, Dataset Card Contact – can be filled in later as more information becomes available. The initial card provides the core necessary information.)