Measuring the Rarity of Heterogeneous Digital Assets through Computer Vision

Conventional rarity scores rely on the meta-data of an NFT collection to determine relative rarity within a collection. Such data is provided by the developers of the collection. In this paper, we discuss pixel-based rarity, which utilizes on raw pixel data to determine rarity. More specifically, we apply an algorithm that captures the pixels in the NFTs and assigns a rarity score to each pixel. It weighs the scores across all pixels within a set of NFTs to give an overall rarity score. We argue that by extracting the pixel data directly from the picture we could generate a richer and more systematic rarity computation. We compare rarity scores to pixel-based rarity scores and discuss their respective merits. Our results show that both scores, on average, show similar performance in predicting relative prices. A surprising result is that rarity and pixel rarity are weakly correlated. We argue that pixel rarity could be a viable metric in the case that the trait data does not properly reflect the visual features of the NFTs. Additionally, pixel-based rarity can be used to compare NFTs across multiple collections and could be a viable substitute for rarity scores where there is limited or no feature data available to compute conventional rarity scores.