“The iNaturalist Species Classification and Detection Dataset” was presented at CVPR 2018 by Grant Van Horn, Oisin Mac Aodha, Pietro Perona, Serge Belongie and colleagues from Caltech, Cornell Tech, and Google. It turned the photo archive of the iNaturalist citizen-science platform into a deliberately hard image-recognition benchmark built around the messiness of the natural world.
The dataset contains 859,000 images representing over 5,000 species of plants and animals, every image verified by multiple citizen scientists. Unlike earlier benchmarks such as ImageNet, where categories are roughly balanced, iNaturalist mirrors reality: some species are photographed thousands of times and others only a handful, producing a steep long-tailed class imbalance. The images also vary widely in quality, camera type, and setting, and many species are visually almost identical, demanding fine-grained discrimination.
The difficulty showed in the results. The authors reported that the best non-ensemble methods of the day reached only about 67% top-one accuracy, well below the near-saturated scores models achieved on cleaner datasets. Performance collapsed for species with few training examples, which the paper highlighted as a call for better low-shot and long-tail learning. The dataset went on to anchor the annual iNaturalist (iNat) challenge and became a standard testbed for fine-grained recognition and species identification systems.
Why business readers should care: this dataset showed that crowdsourced data from enthusiasts can become world-class research infrastructure, and that the long tail - rare cases with little data - is where most real machine-learning systems struggle, not the common cases that dominate demos.