---
title: Religious Artwork Analysis
emoji: 🕌
colorFrom: yellow
colorTo: purple
sdk: docker
app_port: 7860
short_description: Interactive t-SNE explorer for religious artworks
---

# Religious Artwork Analysis

Code for **"Analysis of Artworks Across Different Religions"** — a study of
whether (and which) visual features separate religious artworks of Buddhism,
Christianity, Hinduism and Islam, using a hand-verified, balanced dataset of
3,997 paintings (~1,000 per religion).

- **Dataset**: [Religious Artwork Dataset on Kaggle](https://www.kaggle.com/datasets/zigaklun/religious-artwork-dataset)
  (3,997 images from 8 museum/archive sources, hand-verified labels, CC BY-NC 4.0)
- **Live demo**: interactive t-SNE explorer (see [`app/`](app/))

## Key result

Semantic features transfer across museums; style features largely do not.
Religion-classification accuracy (chance = 0.25):

| feature family        | dims | pooled 5-fold | leave-one-source-out |
|-----------------------|-----:|--------------:|---------------------:|
| CLIP attribute scores |   27 |         0.919 |             **0.856** |
| CLIP embedding        |  512 |         0.941 |                 0.828 |
| DINOv2 embedding      |  768 |         0.910 |                 0.783 |
| hand-crafted (all)    |  378 |         0.741 |                 0.391 |
| pose (main figure)    |   36 |         0.611 |                 0.369 |
| face composition      |   39 |         0.501 |                 0.280 |

The pooled–LOSO gap measures **source leakage**: features that encode museum
reproduction style (scan texture, framing) look discriminative pooled but
collapse on unseen sources. Reproduce with `evaluation/family_accuracy.py`.

## Setup

```bash
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# 1. data (needs a Kaggle API token)
python data/download.py

# 2. preprocessing: guarded background masks (~1 h, CPU)
python preprocessing/generate_masks.py

# 3. features (each checkpointed & resumable)
python features/extract_handcrafted.py --workers 8
python features/extract_clip.py
python features/extract_dino.py
python features/extract_faces.py
python features/extract_pose.py

# 4. evaluation table
python evaluation/family_accuracy.py

# 5. web app
uvicorn app.main:app --port 8000
```

## Repository layout

```
data/           dataset download + expected layout
preprocessing/  crop_padding + guarded U^2-Net background masking
features/       one extractor per feature family (see features/README.md)
evaluation/     pooled vs leave-one-source-out accuracy per family
app/            FastAPI + Plotly interactive t-SNE explorer
```

## Authors

Timotej Cvikl & Žiga Klun — Faculty of Computer and Information Science, University of
Ljubljana. Code under MIT license; dataset under CC BY-NC 4.0 (see the Kaggle page).