We audited 14,512 heritage-site photos with a vision model. The "trustworthy" sources were the dirtiest.
We run Kahve Tabela , an open atlas of 32,000+ registered heritage sites in Türkiye — castles, ancient cities, mosques, museums. Like everyone building a dataset on a budget, we filled the photo gaps from the usual open sources: Wikipedia article images, Wikimedia Commons geosearch, Google Places, Mapillary, the national heritage inventory. Then a reader reported that a photo on one of our pages…
We conducted a comprehensive audit of 14,512 heritage-site photos, using a local vision model. The study revealed two crucial insights: the quality of the source photos and the limitations of relying on names during the auditing process. The most problematic source was Wikipedia article images, which had a 64.3% error rate, nearly nine times higher than Google Places images.
This discrepancy occurs because articles often contain unrelated content, such as maps or portraits, which the model mistakenly identifies as part of the subject. Additionally, when the model is prompted with the name of a place, it tends to hallucinate geographical information, leading to inaccurate accusations. To improve accuracy, the study recommends deleting photos that fail both passes of the audit, rather than relying solely on the first pass.
Furthermore, it is advised to avoid using the name of the place in the prompt and instead focus on the visual content itself. The full dataset, along with the auditing methods, is available on Zenodo and Kaggle for further exploration and verification.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
