Olfaction is a crucial element of human experience but has not gained a lot of attention in cultural heritage. The ICPR-ODeuropa Olfactory Recognition (ODOR) challenge has been created in the context of the EU-funded Odeuropa Project, which aims to remedy this shortcoming by promoting, preserving, and recreating the olfactory heritage of Europe.

Through this challenge, we want to promote the development of object detection algorithms that work under realistic conditions, such as varying image quality and modalities, long-tailed category distributions, or fine-grained detection classes. Being able to detect olfactory objects (e.g. tobacco pipes, perfumed gloves) might in turn lead to the ability to recognize more complex, implicit smell references such as smell gestures or olfactory iconography. The challenge thus promotes a multisensory perspective in computer vision and digital humanities.

Participants are challenged to create systems to detect a diverse range of smell-related objects in historical artworks. The detection targets are objects that either emit strong smells themselves, like different kinds of flowers, or else implicitly hint at the presence of a smell, e.g. flies. In contrast to standard benchmarking datasets like COCO or ImageNet, the ODOR challenge dataset emerges from an application outside of computer vision itself, which leads to multiple challenging properties that participants have to overcome:

• Object detection in the artistic domain of paintings and prints requires algorithms to cope with varying degrees of abstraction and artistic styles, which leads to an intra-class variance considerably higher than that of photographic depictions.
• Unlike standard object detection datasets, where images usually contain repetitive objects with a large number of instances per sample, historical artworks often contain many object instances of diverse sizes. In addition, these instances are more often occluded by other objects.
• Smell-relevant objects can be very particular and specific, leading to a fine-grained classification of target objects. Different types of flowers for example might have a different smell even though they look very similar.
• Since the dataset covers a time span of multiple centuries, the appearance of some target objects is subject to historical change. Special man-made objects like gloves or beverage vessels might have changed their appearance over the years, whereas others like flowers or animals remained invariant.

The challenge is held via Codalab. Participants can find more details and register for the challenge via CodaLab.

A csv-file containing the class hierarchy and distribution of the 15484 training annotations can be downloaded here.

Starting kits to easily recreate and modify a baseline model for detection can be found on the codalab page or at GitHub.