Big Bird: A global dataset of birds in drone imagery annotated to species level. Version 1.0

Full description

Big Bird is a globally diverse dataset of images to support the development of computer vision tools for detecting and identifying birds in drone survey imagery. The dataset contains 23,865 images spanning broad biological, digital, and environmental contexts, including over 100 species of birds, 21 different camera systems, and all seven continents. Birds in a subset of 4,824 images are annotated with bounding boxes and biological attributes, including species, posture category, age category, and sex. This subset includes 49,990 bird annotations from 101 species. The dataset is intended to facilitate more efficient monitoring methods to inform effective conservation. 1. annotated_dataset.zip: Contains images and COCO format annotations.2. dataset/: Contains the full collection of drone images, including both annotated and unannotated data. The data are provided as multiple zipped subfolders grouped by contributor. Data within each sub folder are then grouped into subdirectories by survey. Within each survey directory, images are divided into labelled and unlabelled directories. Labelled images are further split into four directories, one containing images without birds, and three containing images labelled with boxes, polygons, or a mix of both.3. testtrain_dataset.zip: Contains the test and train subsets used in the associated paper.4. supporting_information.zip: Contains supporting documentation, demo images, model, supporting tables, code, and environments to assist users in recreating the analysis undertaken in the associated paper. See the README in the supporting information folder for more info.This dataset is intended for: - Training and evaluating computer vision models for bird detection and identification in drone imagery - Benchmarking model performance across diverse ecological contexts - Developing automated methods for drone-based bird surveys - Supporting research in ecological monitoring and conservation management. The dataset is not intended to replace field surveys or expert ecological interpretation, but to support scalable and efficient analysis of drone imagery.