mingbohong/Bridge-Diverse-Weather-DroneVehicle

Diverse Weather DroneVehicle

This dataset repository releases COCO-style annotation files for Diverse Weather DroneVehicle, a UAV-based domain-generalization object detection benchmark built from the original VisDrone/DroneVehicle dataset. The benchmark organizes images into four weather and illumination domains: day, dark, extreme_dark, and foggy.

The default protocol trains detectors on the clear daytime source domain and evaluates them on unseen adverse target domains. This setting tests robustness to realistic UAV perception shifts, including fog, low light, extreme darkness, small objects, dense traffic, and changes in object scale.

Files

File	Domain	Role
day.json	Clear/day	Source training domain
dark.json	Dark	Target test domain
extreme_dark.json	Extreme dark	Target test domain
foggy.json	Foggy	Target test domain

All four files follow the COCO detection annotation format and contain the same five categories: car, truck, freight_car, bus, and van.

Weather Domains

The benchmark includes clear daytime, dark, foggy, and extreme-dark UAV scenes. The bottom row visualizes enhanced extreme-dark examples for illustration only; the benchmark evaluates on the original target-domain images.

RGB Histogram Analysis

Average RGB histograms reveal strong appearance gaps across domains. Clear images have a balanced mid-intensity distribution. Dark scenes shift toward low pixel values. Foggy images concentrate around high-intensity, low-contrast regions. Extreme-dark images peak sharply near zero, with average RGB values clustered around 15, indicating severe illumination loss.

Annotation Statistics

Weather	Images	BBoxes	Area mean +/- std	Car	Truck	Freight car	Bus	Van
Clear/day	8,881	132,942	2,830 +/- 3,597	110,815	11,016	4,131	4,408	2,572
Dark	13,553	207,535	2,493 +/- 2,504	183,059	6,190	5,061	6,905	6,320
Extreme dark	4,965	85,006	3,079 +/- 3,329	72,955	3,750	3,236	3,466	1,599
Foggy	1,040	27,087	1,408 +/- 1,646	22,950	1,167	972	554	1,444
Total	28,439	452,570	-	389,779	22,123	13,400	15,333	11,935

Foggy scenes contain fewer annotated objects and smaller average bounding boxes than the clear, dark, and extreme-dark domains. The benchmark therefore tests more than color or illumination transfer: models must also handle inconsistent object scales and visibility across weather conditions.

Recommended Data Layout

Place the original DroneVehicle image folders and these annotation files under the same root:

data/
  DroneVehicle/
    day/
    dark/
    extreme_dark/
    foggy/
    day.json
    dark.json
    extreme_dark.json
    foggy.json

For MMDetection-style configs, set ann_file to the corresponding JSON file and data_prefix to the matching image directory.

Benchmark Protocol

The standard domain-generalization split is:

1. Train on day.json with images from day/.
2. Evaluate separately on dark.json, extreme_dark.json, and foggy.json.
3. Report COCO-style mAP for each target domain and their average.

This protocol is used in:

Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization
CVPR 2026

Citation

@inproceedings{hong2026bridge,
  title={Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization},
  author={Hong, Mingbo and Liu, Feng and Gevaert, Caroline and Vosselman, George and Cheng, Hao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

Notes

• This repository contains annotation files and dataset documentation.
• The source images come from the original VisDrone/DroneVehicle dataset.
• Please follow the original DroneVehicle terms when using the image data.

Original Dataset

If you use the image data, please also cite and follow the terms of the original DroneVehicle dataset:

@article{sun2020drone,
  title={Drone-based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning},
  author={Sun, Yiming and Cao, Bing and Zhu, Pengfei and Hu, Qinghua},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2022},
  pages={1-1},
  doi={10.1109/TCSVT.2022.3168279}
}