🌍 Dataset Card for Real-World Distribution Shifts (RWDS)

This repository contains the data presented in Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery.

📊 Dataset Description

• Homepage: https://RWGAI.com/RWDS/
• Repository: https://github.com/RWGAI/RWDS
• Paper: Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery
• Leaderboard: N/A
• Point of Contact: [email protected]

🎯 Dataset Summary

The Real-World Distribution Shifts (RWDS) dataset is a suite of three novel domain generalisation benchmarking datasets that focus on humanitarian and climate change applications. These datasets enable the investigation of spatial domain shifts in satellite imagery-based object detection under real-world scenarios.

RWDS addresses the lack of standardised benchmark datasets for assessing object detection under realistic domain generalisation scenarios. The datasets evaluate model robustness when target distributions differ from source data, particularly focusing on disaster assessment and climate applications.

🏆 Supported Tasks and Leaderboards

Primary Task: Object Detection with domain generalisation

• object-detection: The dataset can be used to train object detection models and evaluate them under domain shifts in satellite imagery.
• domain-generalisation: Evaluate model performance across different climate zones and various disaster scenarios using two experimental setups:
  • Single-source setup: Train on one source domain and evaluate on unseen target domains respectively.
  • Multi-source setup: Train on multiple source domains and evaluate generalisation to held-out target domain.

🗣️ Languages

The dataset contains satellite imagery with English annotations and metadata.

🏗️ Dataset Structure

📈 Data Instances

Each data instance contains and falls within a specific domain in RWDS-CZ, RWDS-FR and RWDS-HE:

• image: A satellite image (512x512 pixels)
• objects: Bounding box annotations with class labels

🗂️ Data Fields

RWDS-CZ (Climate Zones)

• image: PIL Image object
• objects: List of bounding boxes with 16 object classes
• Domain: One of ["CZ_A", "CZ_B", "CZ_C"] (Tropical, Arid, Temperate)
• Task: Object detection and identification in satellite imagery.

RWDS-FR (Flooded Regions)

• image: PIL Image object
• objects: List of bounding boxes with binary damage classification
• Domain: One of ["US", "India"]
• Task: Detecting damaged buildings in flooded regions.

RWDS-HE (Hurricane Events)

• image: PIL Image object
• objects: List of bounding boxes with binary damage classification
• Domain: One of ["Florence", "Michael", "Harvey", "Matthew"]
• Task: Detecting damaged buildings across different hurricane events.

📊 Data Splits

The dataset is split into training, validation, and test sets for each domain to enable proper domain generalisation evaluation using leave-one-domain-out methodology.

🔨 Dataset Creation

📚 Curation Rationale

The dataset was created to address the critical need for robust object detection models in damage assessment and climate change applications, where models must perform well across different geographic regions and environmental conditions without access to target domain data during training.

🗃️ Source Data

Initial Data Collection and Normalisation

RWDS is derived from two established satellite imagery datasets that have been thoroughly preprocessed and reorganized for domain generalisation research:

• RWDS-CZ utilizes the xView dataset, which provides high-resolution (0.3m) satellite imagery captured at a global scale across 60 object classes
• RWDS-FR and RWDS-HE are built from the xBD (xView Building Damage) dataset, which contains satellite imagery focused on building damage assessment from various natural disasters

Both source datasets provide geo-coordinates that enable the creation of domain-specific splits based on geographic and climatic factors. The original datasets have been extensively processed, including class filtering, domain mapping, and data splitting procedures to create meaningful domain generalisation benchmarks.

🏷️ Annotations

Annotation process

• RWDS-CZ: Object classes filtered to include only those present across all climate zones (minimum 30 samples per class)
• RWDS-FR/HE: Original multi-class damage annotations converted to binary classification (Damaged vs No Damage)
• Domain mapping based on geo-coordinates and climate/event classification
• Systematic data splitting using Algorithm 1 from the paper to ensure balanced distributions

Who are the annotators?

Original annotations from xView and xBD datasets, processed and reorganised by the RWDS dataset creators.

🤔 Considerations for Using the Data

🏷️ Social Impact of Dataset

This dataset supports humanitarian applications and climate change research by enabling the development of more robust object detection models for disaster response and environmental monitoring.

📖 Additional Information

📄 Dataset Curators

Sara A. Al-Emadi, Yin Yang, Ferda Ofli (Qatar Computing Research Institute & College of Science and Engineering, Hamad Bin Khalifa University)

📚 Licensing Information

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0), similar to the licensing of the source datasets:

• xView dataset license for RWDS-CZ
• xBD dataset license for RWDS-FR and RWDS-HE

📝 Citation Information

@inproceedings{alemadi2025rwds,
  title={Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery},
  author={Al-Emadi, Sara A. and Yang, Yin and Ofli, Ferda},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}