iNaturalist 2018 Species Classification Competition

By Oisin Mac Aodha and Grant Van Horn

Computer vision will play a crucial role in visual search, self-driving cars, medicine and many other applications. Success will hinge on collecting and labeling large labeled datasets which will be used to train and test new algorithms.

One area that has seen great advances over the last five years is image classification i.e. determining automatically what objects are present in an image. Existing image classification datasets have an equal number of images for each class. However, the real world is long tailed: only a small percentage of classes are likely to be observed; most classes are infrequent or rare. As an example, photographs of plant and animal species are heavily imbalanced: a few species are abundant, while most species are rare.

It is estimated that the natural world contains several million species. Accurately monitoring species in the wild is an important step for measuring the health of the environment. However, most species will occur only rarely and experts will only recognize a few of them. This makes automatic species classification in images a really interesting test case for computer vision.

To stimulate progress in this area, Caltech researchers from the Visipedia project and the iNaturalist team have created the 2018 iNaturalist Image Classification Competition (iNat2018). iNat2018 contains over 450 thousand training images from 8,142 different species of birds, mammals, reptiles, plants, among others. The data has been collected and annotated by citizen scientists on the iNaturalist platform. In last year’s competition the winning entries’ performance was very encouraging for species that had lots of training data; however, rare species presented a challenge. Machines were less data-efficient as compared to humans, and species with fewer than a hundred training examples were particularly challenging to learn for computer vision algorithms. This year’s dataset is even more challenging as the number of species has been increased, resulting in more species with fewer training images.

Compared to iNat2017, the training distribution of iNat2018 has an even longer tail, where the vast majority of species have fewer than 100 training examples each. Some example species are shown on the right.

You are invited to take part in the competition on Kaggle, with the final submission due in June 2018. Winners will be announced at the Fifth Workshop on Fine-Grained Visual Categorization (FGVC5) co-organized by BYU, Caltech, Cornell, Google, Microsoft AI for Earth, MIT, and UMass at CVPR 2018. Training data, annotations, and links to pretrained models are available here. Best of luck!