ArtPlaces

Olfaction is often overlooked in cultural heritage studies, while examining historical depictions of olfactory scenes can offer valuable insights into the role of smells in history. The main challenge is the lack of published datasets with scene annotations for historical artworks, especially in artistic fragrant spaces.

We introduce a novel artistic scene-centric dataset, called ArtPlaces, consisting of 4541 artworks and categorized across 170 distinct physical scene categories. Additionally, it features two test sets aimed at the classification of fragrant places and artistic scenes. ArtPlaces is derived from three source datasets, RASD and WASD established by retrieving images from the Rijksmuseum collection[1] and Wikidata[2]. The respective query terms are used as weak (ie. semi-automatically generated) labels which serve as supervision signals during fine-tuning. The third source dataset is Fragrant Spaces, a manually labeled dataset of olfaction-related artworks. All the data are annotated with the Places365 categories[3]. 

We show that a transfer-learning approach using weakly labeled training data can remarkably improve the classification of fragrant spaces and, more generally, artistic scene depictions. This work lays a foundation for further exploration of olfactory spaces recognition and broadens the classification of physical scenes to the realm of fine art.

References

[1] Rijksmuseum. (2024). Object Metadata API. Retrieved from https://data.rijksmuseum.nl/object-metadata/

[2] Wikidata. (n.d.). Query Service. Retrieved from https://query.wikidata.org/

[3] CSAILVision. (n.d.). Places365. GitHub. Retrieved from https://github.com/CSAILVision/places365/tree/master?tab=readme-ov-file