Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning

Norouzzadeh and colleagues demonstrated that deep convolutional neural networks could automate the labor-intensive task of processing millions of camera-trap images. Training on the labeled Snapshot Serengeti dataset, their models identified species, counted individuals, and described behaviors with accuracy approaching human volunteers for the most common species, while flagging uncertain images for human review. The system could process images vastly faster than manual annotation, projecting enormous savings in expert time. This landmark proof-of-concept showed that machine learning had matured enough to handle the camera-trap data deluge, catalyzing tools like MegaDetector and Wildlife Insights. It is among the most cited papers bridging artificial intelligence and wildlife ecology and reshaped expectations for what automated image analysis could deliver to conservation.