Many research groups and governmental organizations, including the US Environmental Protection Agency, collect samples of insect larvae from freshwater streams to assess the health of stream ecosystems. Specimens in the samples are manually identified to the level of species or species group and counted. This is expensive, time-consuming, and requires many years of experience. Automating this visual identification task requires highly-accurate fine-grained recognition methods. This abstract describes three databases of high-resolution images created to promote the development of such methods, presents benchmark results on these images, and discusses some of the issues raised for fine-grained recognition.