Spaces:
Sleeping
Apply for a GPU community grant: Academic project
Religious Artwork Analysis — interactive explorer for a peer-reviewed academic study
This Space hosts the interactive companion app for our paper "Cross-Religious Artwork Analysis Using Computer Vision and Deep Features" (University of Ljubljana, Faculty of Computer and Information Science, to appear at the ERK conference). The project studies whether visual features separate religious artworks of Buddhism, Christianity, Hinduism, and Islam — and which of those signals are genuine iconography versus museum-digitization artifacts.
Everything is open: the hand-verified, balanced dataset of ~4,000 artworks from eight open-access museum collections is published on Kaggle (CC BY-NC 4.0), the code is MIT-licensed on GitHub, and this Space lets anyone — including art historians without programming skills — explore the collection on a joint t-SNE canvas, reweight feature groups (hand-crafted, CLIP, DINOv2, pose, faces) in real time, edit CLIP attribute prompts to test their own iconographic hypotheses, and run automatic cluster detection with human-readable labels.
The app is CPU-bound: on-demand t-SNE reprojection over 4,000 images × ~1,800 feature dimensions takes minutes on the free 2-vCPU tier, which makes live exploration frustrating for visitors. A CPU upgrade tier (8 vCPU / 32 GB) would make interactive recomputes fast enough for real use in teaching and demos. No GPU is needed — all embeddings are precomputed.
Dataset: https://www.kaggle.com/datasets/zigaklun/religious-artwork-dataset
Source code: https://github.com/cvikl/religious-artwork-analysis