google/cameratrapai

What it solves

Manual processing of images from motion-triggered wildlife cameras (camera traps) is a significant bottleneck for conservationists. SpeciesNet automates this by identifying and classifying wildlife species in these images, allowing practitioners to focus more on conservation and less on manual image review.

How it works

SpeciesNet uses an ensemble of two AI models:

1. Object Detector: Uses MegaDetector to locate animals, humans, and vehicles within an image.
2. Image Classifier: Uses a model trained on over 65 million images with an EfficientNet V2 M architecture to classify the detected objects into one of more than 2,000 labels (including species, higher-level taxa, and non-animal classes).

The ensemble combines these models using heuristics and optional geographic information (such as country or state codes) to filter predictions and assign a single category to each image.

Who it’s for

Conservation practitioners and wildlife monitoring researchers who need to process large volumes of camera trap imagery.

Highlights

• Large-scale training: Trained on a geographically diverse dataset of 65M+ images.
• Broad classification: Supports over 2,000 labels covering diverse species and taxa.
• Geographic filtering: Can use ISO country codes and US state abbreviations to improve prediction accuracy.
• Flexible deployment: Available as a Python package, through cloud platforms like Wildlife Insights and Animl, or via graphical tools like AddaxAI.