UAV Wildlife Detection Datasets: Kangaroo, Koala, and Waterbird

1. Included DatasetsThe Kangaroo dataset consists of thermal images captured at a resolution of 640×512 using drone-mounted sensors. It includes 1,149 training images and 289 validation images. The annotations cover three classes: Kangaroo, Non-Arboreal (representing canopy-level heat sources), and Non-Kangaroo (unidentified ground-level objects). Annotations are provided in both YOLO .txt format and COCO-style .json files.
The Koala dataset also uses thermal imagery with a resolution of 640×512 and was collected in dense eucalyptus forest environments. The dataset comprises 4,277 training images and 1,252 validation images, annotated with two classes: Koala and Non-Koala. Similar to the Kangaroo dataset, annotations are available in both YOLO and COCO formats, enabling compatibility with a wide range of object detection frameworks.
The Waterbird dataset contains high-resolution RGB imagery (5472×3648 and 8192×5460) collected in wetland environments. It includes 390 training images and 128 validation images. This dataset is annotated with ten fine-grained bird categories, including Eurasian Coot, Chestnut Teal, Grey Teal, and others. Annotations are provided in both YOLO and COCO formats, supporting multi-class and small object detection in complex natural scenes.
2. Folder Structure
Each dataset directory follows this structure:
[Dataset Name]/
├── images/ # UAV-captured image frames
├── annotations_yolo/ # YOLO-format annotations
├── annotations_coco/ # COCO-style JSON annotations
├── README_[Dataset].md # Dataset-specific documentation
3. Annotation Formats
YOLO Format:
[class_id x_center y_center width height] — normalized coordinates per object.
COCO Format:
images, annotations, and categories fields, compatible with standard tools like pycocotools, MMDetection, and Detectron2.
4. Use Cases
Small object detection in thermal and RGB drone imagery
Ecological monitoring and conservation AI
Multi-class and fine-grained object recognition
Evaluation of domain adaptation and model generalization under real-world UAV scenarios