A stereo dataset of annotated budgerigar flight trajectories for multi-agent collision avoidance studies

This article presents a collection of 360 synchronized stereo video pairs capturing the flight behaviour of Budgerigars (Melopsittacus undulatus) in a controlled indoor arena. The recordings were acquired at 120 frames per second using a calibrated stereo camera setup with a fixed baseline. The dataset includes both solo flight sequences and structured head-on interaction scenarios involving one-to-one, two-to-two, and three-to-three group configurations. All individuals are manually annotated in both camera views using identity-consistent bounding boxes along with four body keypoints: head, tail, left wing, and right wing. In total, the dataset contains 2760,178 labelled annotations. The synchronized two-dimensional annotations from both views are reconstructed into metric three-dimensional coordinates using calibrated stereo triangulation, providing frame-level trajectories and pose information. The dataset includes raw stereo video files, annotation files in CVAT XML format, processed three-dimensional trajectory data in CSV format, stereo calibration parameters, and scripts for reconstruction. Technical validation measures are provided to document dataset quality, including stereo calibration consistency assessed using stereo reprojection error of 0.50 pixels, annotation reliability analysis, and qualitative assessment of physically plausible motion patterns. By integrating synchronized stereo recordings, identity-consistent 2D annotations, and reconstructed 3D trajectories within a single resource, the dataset supports applications such as multi-object tracking, 3D pose estimation, trajectory modelling, and the analysis of multi-agent interaction, while detailed wing kinematic analysis is constrained by interpolation uncertainty and rolling-shutter effects.