Datasets

This page introduces how to download and use our released datasets.

Introduction

We introduce two new real-world aerial view datasets, LINZ and UGRC, captured in Selwyn (New Zealand) and Utah (USA), respectively. Both datasets have ground sampling distance (GSD) of 12.5 cm per px and have been sampled to 112 px × 112 px image size. For data annotation, we label only the small vehicle centers. To leverage the abundance of bounding box-based open-source object detection frameworks, we define a fixed-size ground truth bounding box of 42.36 px × 42.36 px center at each vehicle. Annotations are provided in COCO format [x, y, w, h], where "small" in the annotation json files denotes the small vehicle class and (x, y) denotes the top-left corner of the bounding box. We use AP50 as evaluation metrics. For more details, please check our paper.

Download

You can download LINZ and UGRC using this command:

wget https://datastore.shannon.humansensing.cs.cmu.edu/api/public/dl/IbKPrtJW -O aerial_vehicle_detection_dataset.zip

The zip file contains two parts. The "real" folder contains Real LINZ and UGRC data, while the "synthetic" folder contains the synthetic data we generate based on the steps here and annotated based on the steps here. After unzipping the data, please organize the "real" folder in the following format:

|-- Data
    |-- Real
        |-- LINZ
            |-- test
                |-- images
                    |-- 0001_0001_0000001.jpg
                    |-- ...
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
            |-- train
                |-- images
                    |-- 0002_0001_0000001.jpg
                    |-- ...
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
            |-- validation
                |-- images
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
        
        |-- UGRC
            |-- test
                |-- images
                    |-- 12SVK260260_0000001.jpg
                    |-- ...
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
            |-- train
                |-- images
                    |-- 12TVK100560_0000001.jpg
                    |-- ...
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
            |-- validation
                |-- images
                |-- annotations_coco_FakeBBoxes:42.36px_ForIoU:0.500.json
        

We provide the checkpoints of four detectors trained on LINZ and UGRC.

Inference

We provide several examples of using open-set models to directly evaluate performance on our datasets.
inference: Examples using Gemini, InternVL3, and DeepSeek-VL2 to inference on UGRC test set.
ConvertPseudoAnn.py: Convert predicted bounding boxes to pseudo annotations.
EvaluatePseudoAnn.py: Compute precision and recall between ground truth and predicted results.