Advances in Digital Music Iconography. Benchmarking musical instrument localization in non-photorealistic images

We present Minerva (‘Musical INstrumEnts Represented in the Visual Arts’): a novel benchmark data set in the field of object detection, focused on the detection of musical instruments in non-photorealistic image collections. This task can be situated in the field of music iconography, a discipline on the brink of musicology and art history, studying the object, themes, and subject matter relating to music as they are represented in the visual arts. Our benchmark experiments highlight the feasibility of the classification and detection tasks under scrutiny but also, and perhaps primarily, the significant challenges that state-of-the-art machine learning systems are still confronted with on this data.