axel-riben/arcdataset-brutalism-extension

Architectural Styles Dataset (Curated and Extended)

Dataset Summary

A curated and extended version of dumitrux's Architectural Styles Dataset. The original dataset covered 25 architectural styles; 630 images were removed by automated filters (duplicates, low-resolution), leaving 9,483 images. A 26th class, Brutalism, was added from 284 manually curated Wikimedia Commons photographs, bringing the total to 9,767 images across 26 classes.

Intended use: training and evaluating image classifiers that predict the architectural style of buildings from exterior photographs.

Supported Tasks

• —Image Classification (multi-class-image-classification): predict one of 26 architectural style labels from a building photograph.

Dataset Structure

Data Instances

Each instance is a JPEG or PNG photograph of a building, assigned a single architectural style label. Images vary in resolution (minimum 224 px on the short side) and aspect ratio.

Data Fields

Classes

26 architectural styles:

Data Splits

Stratified 70/15/15 train/val/test split, fixed seed 42, applied jointly across all 26 classes.

Dataset Creation

Curation Rationale

The original Kaggle dataset contained duplicates and low-resolution images unsuitable for training CLIP-based classifiers (minimum input size 224 px). Brutalism was added because it is a visually distinctive style absent from the original 25 classes, and high-quality labelled images were available from Wikimedia Commons.

Source Data

Original 25 classes:

• —Danci, Marian Dumitru/dumitrux. (n.d.). Architectural Styles Dataset [Data set]. Kaggle. https://www.kaggle.com/datasets/dumitrux/architectural-styles-dataset
• —Downloaded: 2026-05-05 · Curated: 2026-05-07

Brutalism class:

• —Source: Wikimedia Commons, MediaWiki API
• —Scraped across 40 country-level categories (e.g. "Brutalist architecture in the United Kingdom") with subcategory recursion up to depth 2
• —Scraped & downloaded: 2026-05-08 · Curated: 2026-05-09

Curation Process — Original 25 Classes

• —Automated filters: corrupt files; minimum dimension < 224 px; perceptual-hash duplicates (Hamming distance < 5)
• —Manual review: stratified random sample of 10 images per class; no further images removed
• —Result: 630 images removed from the original 10,113, leaving 9,483 images

Curation Process — Brutalism Class

All 950 downloaded images were reviewed individually. An image was kept only if it showed the exterior of a Brutalist building as the primary subject, was a real photograph (not a render or drawing), had no severe obstructions, and had sufficient resolution and framing for style classification.

Rejection breakdown:

"other" covers heavily post-processed images, composites, and signage-dominated shots.

Leakage Remediation

Nearest-neighbour analysis on CLIP ViT-B/32 embeddings identified building-level duplicates split across train and test in the Ancient Egyptian and Novelty classes. Splits were regenerated after remediation; reported F1 for these classes may still be slightly optimistic.

Considerations for Using the Data

Geographic and Temporal Bias

• —The original 25 classes are heavily Western and European; the Brutalism extension spans 40 countries but remains biased toward countries with strong Wikimedia Commons representation.
• —The dataset is predominantly pre-1980. Parametricism and Metabolism are absent.

Known Limitations

• —Manual review of the original 25 classes was sample-based (10 images per class), not exhaustive.
• —The Brutalism class (284 images) is smaller than most original classes; performance on this class may be sensitive to distribution shift.
• —The model is trained on exterior photographs; performance on interiors, renders, and architectural drawings is undefined.

Sensitive Content

No personally identifiable information. Images are of buildings, not people.

Additional Information

Licensing

• —Original 25 classes: CC0 1.0 (Public Domain)
• —Brutalism class: mixed free licences (CC0, CC-BY, CC-BY-SA) inherited from Wikimedia Commons contributors. Per-image licence information is recorded in the accompanying `brutalism_downloaded.jsonl` manifest. Downstream use must comply with the applicable per-image licence.

Citation

If you use the original Kaggle dataset, please cite:

Danci, Marian Dumitru/dumitrux. (n.d.). Architectural Styles Dataset [Data set]. Kaggle.
https://www.kaggle.com/datasets/dumitrux/architectural-styles-dataset