Can a computer identify bird species in photos? Researchers at Cornell Tech and Caltech have partnered with the Cornell Lab of Ornithology to train Merlin Bird Photo ID to recognize 650 of North America’s most common bird species based on images.
Sometimes Merlin gets it right, just like magic. Other times, Merlin gets it wrong–sometimes way wrong. What’s going on behind the scenes–and how can you help us keep improving Merlin’s accuracy?
The challenge: Asking computers to identify bird species is a challenge not only because some species look so alike, but also because their shape varies from moment to moment. On top of that, photographs of birds often include complex backgrounds, and the birds may be far away or blurry.
The solution: Computer vision researchers create “convolutional network” systems that use patterns in data to train the computer and improve its performance. These systems require massive numbers of images as well as accurate image labels such the type of object, and where the object is in the image. Fortunately, bird watchers are renowned for taking lots of photos and for contributing millions of observations in citizen-science projects. Thousands of people have contributed photos and tagged them to teach Merlin to recognize birds.
In the coming months, we’ll continue to improve Merlin’s accuracy and train it to identify additional species. We’ll also push Merlin’s capabilities to see how well it can identify birds in poorer quality photos taken with mobile devices. Ultimately, we hope that Merlin will be a useful tool to identify birds from around the world–and that its technology will be used to identify other wildlife, plants, and objects too.
We invite you to help us build Merlin Bird Photo ID. Join our email list to receive updates about how you can help–and be the first to hear when we release exciting new features. To support our efforts to develop new technologies to help people understand and enjoy birds, please make a donation. Thank you!