0xgr3y/arch-building-dataset

World Architectural Buildings Dataset (FGIC) for Multi‑Class Image Classification

Multi‑Class Image Classification dataset of world architectural buildings with finalized curation.

Classes

Class	Count	Description
barn	1,680	Traditional wooden barn architecture — residential and storage buildings
bridge	1,680	Various bridge architectures (suspension, arch, truss)
castle	1,680	Medieval and modern castle structures
mosque	1,680	Islamic mosque architecture (domes, minarets)
skyscraper	1,680	Tall commercial and residential, city skyline buildings
stadium	1,680	Sports arenas, stadiums, amphitheaters
temple	1,680	Religious temple structures (Buddhist, Hindu, Asian)
windmill	1,680	Windmill structures (traditional, tower, Dutch)
Total	13,440	All datasets are finalized for curation

Data Collection & Processing

Source

All images sourced from Pexels.com — free stock photos licensed under the Pexels License (free for commercial use, no attribution required).

Collection Pipeline

Images were collected using a multi-mode scraping cascade:

1. Pexels API Key — primary source, direct API access with keyword-based queries per architectural class
2. BrightData — supplementary collection for underrepresented subcategories
3. Selenium — fallback for specific queries requiring browser automation

Search Keywords

Class	Keywords
🏚️ Barn	old wooden barn, barn homes, barn architecture, red barn
🌉 Bridge	bridge, bridge architecture, bridge landscape, bridge building, bridge at night
🏰 Castle	castle architecture, medieval castle, castle building, european castle
🕌 Mosque	islamic mosque, mosque architecture, mosque exterior, beautiful mosque
🏙️ Skyscraper	skyscraper architecture, skyscraper building, tall building, city skyline, skyscraper at night
🏟️ Stadium	stadium design, football stadium, arena stadium, stadium at night, stadium architecture
🛕 Temple	temple building, temple architecture, buddhist temple, ancient temple, temple at night
🎡 Windmill	windmill architecture, traditional windmill, tower mill, windmill exterior, dutch windmill

Compression & Format

• Max resolution: 720px on the longest side, min 200px on the shortest side
• Max file size: 200 KB per image
• Format: JPEG, adaptive quality compression (start: 85, step: −5, min: 40) until file ≤ 200 KB

Quality Filters

Images were filtered based on the following criteria:

Filter	Threshold	Purpose
Minimum resolution	150 × 150 px	Remove thumbnails and icons
Aspect ratio	0.28 – 3.5	Remove extreme panoramas and strips
Color standard deviation	≥ 10.0	Remove near-monochrome images
Minimum file size	5 KB	Remove corrupt/stub files
Maximum file size	4,096 KB	Remove uncompressed originals
File extension filter	Skip .svg, .gif, .webm, .mp4, .ico	Ensure JPEG-only dataset
URL pattern filter	Skip logos, avatars, icons, AI art, illustrations, people, badges, buttons, emojis, text overlays	Remove non-photographic content

Deduplication

Method	Threshold	Description	Artifact
pHash (perceptual hash, 16×16)	Distance ≤ 8	Remove near-duplicate images (visual similarity)	app_pexels.py — implementation; pexels_checkpoint.json — hash records
SHA256	Exact match	Remove byte-identical duplicates	app_pexels.py — implementation; pexels_checkpoint.json — hash records
Human Curation	Full review	Final visual annotation and selection by domain experts	—
Search Keywords	Per-class	Keyword-driven collection with class-specific queries (e.g., "mosque architecture", "dutch windmill")	config_images.json — keyword config; app_pexels.py — scraper script
Collection Checkpoint	Per-batch	Resume-capable checkpoint tracking per class — prevents re-downloading already-collected images across sessions	pexels_checkpoint.json — checkpoint data; app_pexels.py — checkpoint logic

Resolution Distribution

Dimension	Range
Width	256 – 889 px
Height	160 – 1140 px

How to Load

from datasets import load_dataset

dataset = load_dataset("0xgr3y/arch-building-dataset", data_dir="dataset")

Dataset Structure

dataset/
├── barn/         barn_00000.jpg – barn_01679.jpg
├── bridge/       bridge_00000.jpg – bridge_01679.jpg
├── castle/       castle_00000.jpg – castle_01679.jpg
├── mosque/       mosque_00000.jpg – mosque_01679.jpg
├── skyscraper/   skyscraper_00000.jpg – skyscraper_01679.jpg
├── stadium/      stadium_00000.jpg – stadium_01679.jpg
├── temple/       temple_00000.jpg – temple_01679.jpg
└── windmill/     windmill_00000.jpg – windmill_01679.jpg

• Naming convention: {class}_{5-digit-sequence}.jpg (sequential, zero-padded)
• File format: JPEG (all images)
• Total files: 13,440

Dataset Splits

The recommended split is 80% / 10% / 10% using split-folders with seed=42:

Split	Images	Per Class
Train	10,752	1,344
Validation	1,344	168
Test	1,344	168
Total	13,440	1,680

Dataset Comparison

This dataset fills a gap in architectural image classification benchmarks. Existing datasets focus on broader categories:

Dataset	Classes	Images	Domain	Limitation
ImageNet	1000	1.2M	General objects	Few architectural classes, not FGIC-focused
Places365	365	1.8M	Scene recognition	Scene-level, not building-type-level
Oxford Buildings	5K	5K	Oxford-specific	Single city, limited diversity
This dataset	8	13,440	Global architecture	Balanced, multi-cultural, FGIC-focused

Unlike ImageNet (which mixes buildings into "castle", "church", etc.) or Places365 (which classifies scenes, not building types), this dataset provides fine-grained architectural type classification with balanced classes covering diverse global architecture (European castles, Islamic mosques, Asian temples, Dutch windmills, etc.).

Class Selection Rationale

The 8 classes were selected based on:

1. Architectural distinctiveness — each class has visually discriminative features (domes for mosques, verticality for skyscrapers, blades for windmills)
2. Cultural diversity — spans European (barn, castle), Islamic (mosque), Asian (temple), modern (skyscraper, stadium), and utilitarian (bridge, windmill) architecture
3. FGIC challenge level — some pairs (temple/mosque, castle/stadium) share visual features, providing inter-class confusion for evaluation
4. Data availability — sufficient images on Pexels to achieve balanced 1,680 images per class

Limitations & Sampling Bias

• Pexels platform bias: Images sourced from Pexels, a Western-centric photography platform. This may introduce bias toward Euro-American architectural styles and professional photography aesthetics. Users should evaluate on geographically diverse test sets before deployment.
• No inter-annotator agreement metric: Labels were assigned by search keywords (e.g., "mosque architecture") and verified by single-curator human review. Unlike multi-annotator datasets with Cohen's Kappa, this dataset uses keyword-driven collection with human curation, which may introduce curator bias.
• Resolution cap: Images compressed to max 720px longest side. Fine-grained architectural details (e.g., window patterns, ornamental carvings) may be lost at this resolution.
• Single source: All images from Pexels. No cross-platform validation (e.g., Unsplash, Flickr) to assess source-independent classification accuracy.

Ethical Considerations

• No personally identifiable information (PII): All images are photographs of architectural buildings — no faces, names, or personal data are included.
• Human content filtered: URL pattern filters explicitly exclude images tagged with people, portraits, or human-focused content during collection.
• Commercial-free licensing: All images sourced from Pexels under the Pexels License (free for commercial use, no attribution required).
• Building-only scope: This dataset is strictly limited to architectural structures. It does not include content that could be used for surveillance, profiling, or any harmful purpose.
• Responsible use: Users agree not to claim ownership of original images, redistribute without attribution, or use the dataset for unlawful purposes.

Related

• Model: 0xgr3y/Arch-Building-Image-Classification (EfficientNetV2, GeMPooling, Focal Loss, DiscriminativeAdamW LR, SWA)
• Live HF Space: 0xgr3y/arch-building-classifier (Gradio Spaces)
• GitHub Repo: arcxteam/dataset-fgic-architectural (full source code and detail used)

Citation

HuggingFace

@misc{saugani2026_arch_building_dataset_hf,
  title={World Architectural Buildings Dataset (FGIC) for Multi-Class Image Classification},
  author={Saugani},
  year={2026},
  publisher={Hugging Face},
  license={CC-BY-4.0},
  url={https://huggingface.co/datasets/0xgr3y/arch-building-dataset},
  doi={10.57967/hf/9220}
}

GitHub (full source code)

@misc{saugani2026_arch_building_dataset_github,
  title={World Architectural Buildings Dataset (FGIC) for Multi-Class Image Classification},
  author={Saugani},
  year={2026},
  publisher={GitHub},
  license={CC-BY-4.0},
  url={https://github.com/arcxteam/dataset-fgic-architectural}
}