"European Animal Detection Dataset"

"The European Animal Detection Dataset is introduced to support the development and evaluation of Artificial Intelligence (AI)-based methods for wildlife monitoring using camera trap imagery. Existing large-scale datasets, including COCO, ImageNet, MNIST, Pascal VOC, and Google Open Images, were extensively analyzed for training object detection models. However, these datasets were found insufficient for the specific requirements of European wildlife detection due to limited domain relevance, class imbalance, and a lack of species-specific diversity. To address these limitations, a dedicated dataset was constructed using images collected primarily from camera trap sources. The dataset focuses on five classes: Fallow Deer Male, Fallow Deer Female, Fawn, Wild Boar, and Wolf, which are ecologically significant species across Europe. Initially, 10,435 high-resolution images were collected from open-access repositories, including sources informed by the European Food Safety Authority (EFSA) and iNaturalist, ensuring diverse environmental and geographical coverage. Prior to annotation, duplicate and near-duplicate frames were removed using the Structural Similarity Index (SSIM) to improve dataset quality and reduce redundancy. To further enhance model robustness and generalization, extensive data augmentation techniques such as rotation, flipping, translation, and noise injection were applied, expanding the dataset to approximately 20,000 images. All images were manually annotated to ensure high-quality ground truth labels after automated annotation approaches proved inadequate. The dataset includes both daylight and night-vision imagery, capturing real-world variability in illumination and environmental conditions. Annotations are provided in multiple formats, including YOLOv8\/YOLOv10, COCO, and YOLO Darknet, ensuring compatibility with a wide range of deep learning frameworks. The dataset is divided into training (70%), validation (20%), and testing (10%) subsets to support standardized evaluation. This dataset is designed to facilitate research in computer vision\u2013based wildlife monitoring and to benchmark advanced object detection models such as YOLOv8 and YOLOv10. Additionally, it enables the generation of species distribution of heatmaps across Europe, contributing to ecological analysis, wildlife management, and conservation planning. "