README.md

metadata

license: mit
task_categories:
  - image-classification
  - object-detection
tags:
  - carla
  - uav
  - visual-search
  - robotics
  - embodied-ai
pretty_name: CW-AVIS
size_categories:
  - 1K<n<10K

CW-AVIS

CW-AVIS is a CARLA-based benchmark dataset for cross-weather aerial visual instance search. It pairs target descriptions and search start poses with reference images captured under multiple weather and lighting conditions.

The dataset accompanies ReF-AVS.

Contents

• 2,470 PNG reference images
• 61 target definitions across six CARLA towns
• Three search difficulties: easy, medium, and hard
• Clear, cloudy, rainy, wet, night, sunset, and dust-storm conditions
• Total uncompressed size: approximately 2.99 GB (2.78 GiB)

CW-AVIS/
  targets_Town01/
    101_Bus_Stop_in_Front_of_Terrace_Restaurant_by_Bridge/
      target.json
      ref_Clear Noon/
        *.png
      ref_Wet Night/
        *.png
  targets_Town02/
  targets_Town03/
  targets_Town04/
  targets_Town05/
  targets_Town10HD/

Each target.json includes a numeric target ID, target name, natural-language scene hint, ground-truth final XY position, initial UAV poses for each difficulty, and the number of reference images available per weather.

Usage

Download the repository and point ReF-AVS to its root:

export REF_AVS_DATASET_DIR=/path/to/CW-AVIS
python main/run_rivas_batch.py --dataset "$REF_AVS_DATASET_DIR"

SHA256SUMS contains checksums for every PNG and JSON payload file.

License

The dataset metadata and annotations are released under the MIT License. CARLA simulation assets remain subject to their applicable upstream terms.