WorcesterMA_Housing_Facades:

🌐 GitHub | 🤗 Dataset

Street-level photographs of housing facades from Worcester, MA, organized into four facade classes. Each image filename is the property PID (integer). The dataset links housing registry metadata (e.g., year_built) with facade images collected for research in visual housing classification.

Dataset Card

Dataset name: WorcesterMA_Housing_Facades

Short description: Photographs of housing facades from Worcester, MA. Images are labeled into four facade classes (class_1..class_4). Filenames correspond to municipal property IDs (PIDs) and can be joined with the included property metadata.

Size: 22,949 images total

• train: 20,467 images
• test: 2,482 images

Class distribution (train / test):

• class_1: 11,254 / 1,241
• class_2: 4,958 / 543
• class_3: 2,967 / 353
• class_4: 1,288 / 345

Languages: English (metadata fields)

Licensing: license: mit is declared in this repository front matter. Confirm that all source imagery and registry data are cleared for redistribution before publishing externally.

Supported Tasks

• Image classification (housing facade types)

Data Fields / Features

• image — JPEG image file (path under WorcesterMA_Housing_Facades/train or .../test).
• pid — integer property identifier (filename without .jpg).
• location - string housing location address.
• class_label — string label (class_1..class_4) or integer-coded label in registry files.
• split — train or test.
• Additional metadata (where available): year_built, image_path, url in metadata.csv / registry CSVs.

External Generalization Evaluation

We also evaluate model generalization using Google Street View imagery. These images are not included in this dataset due to Google Street View Terms of Service.

However, we provide a Jupyter notebook in the project’s GitHub repository that describes how the evaluation data was collected. Users must ensure compliance with Google’s ToS before running the script.

Recommended Usage

Basic workflow:

1. Verify image integrity (some images may be missing/corrupt).
2. Join images to property metadata by PID if you need additional attributes (e.g., year_built).
3. Use standard image loaders (PIL, OpenCV) and typical transforms for training.

Usage

Load the dataset using the Hugging Face datasets library. Example:

from datasets import load_dataset

# load the 'standard' config (contains metadata + image paths)
ds_train = load_dataset("WorcesterMA_Housing_Facades", "standard", split="train")
ds_test = load_dataset("WorcesterMA_Housing_Facades", "standard", split="test")

print(len(ds_train), len(ds_test))
print(ds_train[0])

Limitations and Biases

• Geographic bias: all images are from Worcester, MA — models trained on this data may not generalize to other cities or architectural styles.
• Class imbalance: class_1 is the largest class; consider class-balancing strategies during training.
• Potential privacy concerns: images may contain visible addresses, license plates, or people. Verify local privacy and usage policies before public release.

Provenance & Source Files

• Image files and shards: WorcesterMA_Housing_Facades/train and .../test folders.
• metadata.csv provides mappings between PIDs and image paths.

Permissions / Access

We have obtained oral consent from the Worcester Assessor’s Office on Dec 15, 2025 to share this dataset. They informed us that all the data available on the website are public records and, therefore, they can be made available by us on a dataset.

Citation

If you use this dataset in a publication, please cite the dataset directory and include any original municipal data sources you used to collect property records and imagery.

About the License

If you believe that this licensing choice is inappropriate, please feel free to contact us. We are open to reconsidering and making changes.

Contact

• Shannon Song: song.shannon.m@gmail.com
• Yiqing Zhang: yiqingzhang125@gmail.com
• Fabricio Murai: fmurai@wpi.edu
• Nancy Ma: nma1@wpi.edu