A large training dataset

The success of our collaboration with our partners has allowed us to gather over a million annotated pictures, likely representing the largest database of this kind in France.

A two-step approach

Our approach is based on two models: 1) a detection model identifies whether an animal/human/vehicle is present in the image, as ’empty’ images are often collected due to false triggers; 2) a classification model predicts to which class (i.e. species or higher taxonomic group) the animal detected belongs to.

Species classified and model performance

Currently, the model uses 41 classes. Each image or video that is submitted, taken by day or by night, is predicted to show either:

• nothing: class ’empty'
• a human presence: classes ‘human’, ‘vehicule’
• an ungulate: classes ‘bison’, ‘chamois’, ‘fallow deer’, ‘ibex’, ‘moose’, ‘mouflon’, ‘red deer’, ‘reindeer’, ‘roe deer’, ‘wild boar’
• a different large mammal: classes ‘bear’, ‘golden jackal’, ’lynx’, ‘wolf’, ‘wolverine’
• a smaller mammal: classes ‘badger’, ‘beaver’, ’lagomorph’, ‘genet’, ‘hedgehog’, ‘marmot’, ‘muskrat’, ‘mustelid’, ’nutria’, ‘otter’, ‘porcupine’, ‘racoon’, ‘racoon dog’, ‘red fox’, ‘squirrel’
• a domestic animal: classes ‘cat’, ‘cow’, ‘dog’, ’equid’, ‘goat’, ‘sheep’
• something else, either ‘micromammal’ or ‘bird’ (which has eight coarse sub-classes).

Images for which the confidence score is below a user-defined threshold are considered ‘undefined’, and should be inspected visually to be classified.

However, for many species predictions are right more than 95% of the time! (results per species here).

Try the software on your own images - and let us know how it performs!
The quality of the predictions improve regularly, as we work on the model and add images to the training database. If you have classified camera-trap pictures or videos, you can help us to improve the model. Contact us to contribute!

You want to learn more about our work?

Rigoudy et al. 2022. The DeepFaune initiative: a collaborative effort towards the automatic identification of the French fauna in camera-trap images. European Journal of Wildlife Research 69:113

Dussert G. et al. 2024. Being confident in confidence scores: calibration in deep learning models for camera trap image sequences. Remote Sensing in Ecology and Evolution 11:88

You could also be interested by:

Dussert et al. 2025. Zero-shot animal behavior classification with vision-language foundation models. Methods in Ecology and Evolution 16:1460.

Dussert et al. 2026. Paying attention to other animal detections improves camera trap classification models. Methods in Ecology and Evolution 17:1248.

The DeepFaune initiative brings together more than a hundred partners, all involved in research, conservation or management of biodiversity: regional or national parks, hunting federations, naturalist associations, research groups and even individuals.
These partners are actively taking part in the project, by contributing camera-trap images or videos to our growing database, and participating in defining the project’s aims and user needs.