AU-AIR

Dataset Overview

Dataset Name

AU‑AIR Dataset

Description

AU‑AIR is a multimodal UAV aerial dataset containing visual data, object annotations, and flight telemetry (time, GPS, altitude, IMU sensor data, velocity) for UAV vision and robotics applications.

Specification

• 8 raw RGB videos (total duration > 2 h)
• 32,283 extracted and annotated frames
• Bounding‑box labels for eight traffic‑related objects: person, car, van, truck, bicycle, motorcycle, bus, trailer
• Each frame provides timestamp, GPS, altitude, IMU sensor data, and linear velocity of the UAV

Download

• Images: link (2.2 GB)
• Annotations: link (3.9 MB)

Output

The dataset is split into training, validation, and test sets, with a separate file for class names.

Conversion Tool

A TFRecord conversion script is provided with the following arguments:

• --annotations: path to annotation file (default ./data/annotations.json)
• --data_dir: directory of image data (default ./data/images/)
• --output_dir: output directory (default ./data/tfrecords)
• --splits: list of split percentages (default 80,10)

Dependencies

The dataset relies on common libraries such as numpy and opencv.