Format: ORAL

Alexander N. Schmidt-Lebuhn1, Pete Thrall2

1 Centre for Australian National Biodiversity Research (a joint venture of Parks Australia and CSIRO), Canberra, Australia 2 CSIRO, Canberra, Australia

Computer vision (CV) has matured as a mobile and user-friendly species identification tool, as demonstrated by apps such as iNaturalist or PlantSnap, which are aimed at the nature-enthusiastic public. However, the CV models of such apps cover taxonomic diversity opportunistically, based on what user-contributed images happened to be available for training the model, as opposed to comprehensive coverage of a taxonomic group. Their scope is primarily the overall appearance of complete, living specimens in their natural environment. If CV is to become the next generation identification tool in taxonomy, field biology, and collection curation, biologists will have to be able to easily train, deploy, and use CV models for anything from dried fruits on a herbarium specimen to pinned insects of their study groups. Specimen collections are increasingly being imaged, and commercial cloud services allow users without expertise in AI research to train models. The main remaining bottleneck is getting species identification models into the hands of end-users. We present Horama ID, a system for the storage and deployment of image classification models for species identification. It consists of a server where contributing taxonomists can upload the model, metadata such as its title and description, who created it, acknowledgements, and a title image, as well as a table of species profiles including profile images and external weblinks, e.g., to eFlora profiles. The Horama app displays the list of models available on the server and allows the user to download them individually and switch between them. The camera page of the app displays live updates of identification estimates. The user can tap suggested names to see species profiles and compare them against the specimen. Horama ID gives taxonomists the opportunity to develop CV models as part of their studies in the confidence that they can become available to end-users.