DeepCamTrap: European Mammal and Bird Recognition in Camera Trap Images

This repository provides models, metric evaluations and the test data sets presented in the papers "Recognizing European mammals and birds in camera trap images using convolutional neural networks" (Schneider et al., 2023) and "Recognition of European mammals and birds in camera trap images using deep neural networks" (Schneider et al., 2024) as well as the dissertation "Attention-based Multitask Learning for Image Analysis" (Schneider, 2026).

We publish annotated camera trap images from two recording locations:

• the Marburg Open Forest (MOF) dataset recorded in Hesse, Germany consisting of about 2,500 images
• the Białowieża National Park (BNP) dataset recorded in Podlaskie Voivodeship, Poland consisting of about 15,000 images

The images are annotated on a 6-level taxonomy (class, group, order, family, genus, species), with the labels grouped into files based on the most precise annotation possible.

The best models from our papers are available in the Tensorflow2 SavedModel format (https://www.tensorflow.org/guide/saved_model). We provide a code snippet to load the models and perform predictions.

File Structure

The repository is structured as follows:

• data contains details about our training data sets as well as a download script for our Marburg Open Forest (MOF) and Białowieża National Park (BNP) test data sets.
• evaluation contains high-resolution images of the confusion matrices from our papers.
• models contains download script for the best models from our 2023 and 2024 papers, the 2026 dissertation and a code snippet to perform predictions with these models.