ResFa: Multimodal Facade Dataset

Overview

ResFa is a multimodal dataset of residential building facades designed for studying structural inference under partial observability. It includes RGB images, infrared thermal images, terrestrial LiDAR point clouds, and hierarchical graph representations.

The dataset supports tasks in multimodal learning, graph neural networks, and building envelope analysis.

---

Dataset Composition

ResFa contains data for 66 residential facades, each including:

• RGB images (multi-view: front, left, right)
• Infrared thermal images (7AM labeled, 3PM unlabeled)
• Terrestrial LiDAR point clouds
• Semantic annotations (window and structural components)
• Hierarchical graph representations (observable and latent nodes)

---

Task Objectives

The dataset is designed for:

• Multimodal facade segmentation
• Structural inference of hidden elements (e.g., studs)
• Graph-based reasoning (observable vs latent structures)
• Cross-modal learning between RGB, IR, and point cloud data

---

Dataset Split

To ensure reproducible benchmarking, the dataset is divided as follows:

• Training set: 40 facades (60%)
• Validation set: 13 facades (20%)
• Test set: 13 facades (20%)

The split is stratified at the facade level to avoid cross-view leakage.

---

Data Format

Each facade contains:

• RGB/ – RGB images (JPEG)
• IR/ – infrared thermal images (JPEG)
• PCD/ – LiDAR point clouds (.pcd)
• graph.json – hierarchical structural graph
• metadata fields (window count, structural estimates)

---

Data Collection

Data was collected using:

• Terrestrial LiDAR scanning systems
• Infrared thermal imaging cameras
• Multi-view RGB photography

All data was collected from residential building exteriors.

---

Annotation

Annotations include:

• Semantic labeling of facade components (windows, walls, structural elements)
• Point cloud classification (coarse and fine labels)
• Graph construction representing structural relationships

Annotations were generated using a combination of manual labeling and computational preprocessing pipelines.

---

Limitations

• Dataset is geographically limited to a single region
• Focused on pre-1975 residential wood-frame buildings
• Infrared data is time-specific (7AM labeled, 3PM unlabeled)
• Structural labels (e.g., studs) are inferred, not directly measured

---

Ethical Considerations

• No human subjects or personal data are included
• All data is collected from publicly visible building exteriors
• Dataset is intended for academic research only
• Not suitable for safety-critical or regulatory applications

---

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

---

Intended Use

This dataset is intended for research in:

• Multimodal machine learning
• Graph neural networks
• Computational building analysis
• Cross-modal structural inference