We are conducting the very first Fieldguide Challenge, a FGVCx challenge focusing on fine-grained classification of Lepidoptera (moths & butterflies). There are an estimated 175,000 species of Lepidoptera (leps), most of which are rarely photographed. Species identification is further complicated by distinct adult and immature stages (caterpillars), and often a lack of concrete visual diagnostics for separation between near-identical species. The goal of the competition is to push state-of-the-art classification of real world data that contains high class imbalance and high infraclass variability.
Moths and butterflies are by far the most photographed group of animals outside of birds.  Having an effective CV component to identification apps would engage tens of thousands of avid citizen scientists. It would allow us to create effective monitoring programs through promotion of night-lighting for moths, which are sensitive environmental indicators. We estimate that at least 50,000 species of moths and butterflies can be identified by images, and we have citizen science groups throughout the world ready to put their smartphone cameras to use.
The Fieldguide Challenge dataset contains over 5000 species, with a combined training and testing set of over 530,000 images that have been collected and verified by multiple authors of Fieldguide. The dataset features visually-similar species, many of which are highly polymorphic, various stages in the species’ lifecycle (larvae vs adult), as well as species that are often misidentified as a moth/butterfly (e.g. caddisflies).
The winning team may have the opportunity to present their results at the FGVC workshop at CVPR in Long Beach. The host would like to work with winning teams to build better models for Fieldguide.ai.
Competition Objectives
Improve classification accuracy over a dataset that contains high class imbalance and high heterogeneity (caterpillar vs. adult, wild vs pinned specimens, etc.)
Improve error detection accuracy - what is not-a-lep*
*Not-a-lep is a single, mixed-bag class containing images commonly mistaken as leps. For example: leafhoppers, caddisflies, data labels, and flowers often included in the background of moth/butterfly photos, etc.