Author: Sara Beery
Biodiversity data is being collected at unprecedented scales, and is no longer able to be efficiently and sustainably processed with human effort alone. Camera traps, motion-activated static cameras used for long-term studies of ecosystems and the wildlife within them, are a prime example of this, collecting hundred of thousands of images per year for each survey. The camera trap community has increasingly turned to computer vision to aide in processing, but the challenges of domain generalization in CV led to the need for in-house ML engineers in order to train usable, reliable model for each project. This doesn’t scale, so we performed a systematic analysis of generalization and sought to deploy a model that would serve the needs of diverse organizations worldwide by increasing human efficiency collaboratively with AI, as opposed to fully automating camera trap data processing. The focus on off-the-shelf generalizability and accessibility led to a a model that has been widely adopted, and is used by over 60 organization and NGOs globally. In this chapter, we discuss the development of the MegaDetector and introduce five diverse end users with different needs and target uses and discuss what made the MegaDetector accessible to them and how it has impacted their conservation and biodiversity work.