SkyFind: A Large-Scale Benchmark Unveiling Referring Expression Comprehension for UAV

SkyFind is the first large-scale dataset tailored for UAV-based Referring Expression Comprehension (UAV-based REC). It provides high-resolution images from Unmanned Aerial Vehicle (UAV) perspectives, bridging the gap between visual localization and natural language understanding. Introduced in our TPAMI 2026 paper, this dataset aims to facilitate human-UAV interaction through natural language.

🌟 Dataset Highlights

• Massive Scale: Features 1,015,638 target-expression pairs, 35,599 images, and 352,910 objects across diverse UAV scenarios.
• Rich Linguistic Diversity: Each original expression in the training set is augmented with two additional variants generated by GPT-4, significantly expanding the vocabulary and sentence structures (tripling the training data volume).
• UAV-Specific Challenges: Specifically designed to address the unique difficulties in aerial scenarios, including abundant background interference, extremely small target sizes ("large scene, small object"), and complex referring relations.
• Multi-Task Applicability: Ideal for Referring Expression Comprehension (REC), multi-object detection and localization, vision-language modeling, and cross-modal retrieval.

📁 Dataset Structure

The dataset is hosted on Hugging Face as partitioned Parquet files (2-5GB shards), ensuring efficient loading and processing. The structure is as follows:

SkyFind/
├─ data/
│  ├─ train/           # Training set parquet shards (331,364 pairs)
│  ├─ validation/      # Validation set parquet shards (5,000 pairs)
│  ├─ test/            # Test set parquet shards (16,546 pairs)
│  └─ train_aug/       # Augmented training set parquet shards (GPT-4 augmented)
├─ Train.json          # Annotations for the original training set
├─ Val.json            # Annotations for the validation set
├─ Test.json           # Annotations for the test set
├─ Train_Aug.json      # Annotations for the augmented training set
├─ bigfile_part_00.zip ... bigfile_part_11.zip  # Zipped raw images
├─ README.md

The annotations field in each Parquet split contains the bounding boxes (bbox) and their corresponding expressions list.

🚀 Quick Start

First, install the required libraries:

pip install datasets huggingface_hub pyarrow

Loading via Hugging Face datasets

You can easily load the dataset using the following Python code:

from datasets import load_dataset

train_data = load_dataset("kunyu241/SkyFind", split="train")
val_data = load_dataset("kunyu241/SkyFind", split="validation")
test_data = load_dataset("kunyu241/SkyFind", split="test")
train_aug_data = load_dataset("kunyu241/SkyFind", split="train_aug")

print(train_data[0])

Example Output:

{
  "fileName": "00001.jpg",
  "image": {"path": "00001.jpg", "bytes": b"..."},
  "annotations": [
    {
      "bbox": [x1, y1, x2, y2],
      "expressions": ["a red car", "the vehicle on the left"]
    }
  ]
}

📥 Downloading Raw Resources (Optional)

If you prefer to work with the raw images and JSON annotations locally:

1. Download the split ZIP chunks (bigfile_part_00.zip to bigfile_part_11.zip) and merge them into a single images.zip file.
2. Download the four JSON annotation files corresponding to each split:
   • Train.json → Matches the train split
   • Val.json → Matches the validation split
   • Test.json → Matches the test split
   • Train_Aug.json → Matches the train_aug split (includes tripled linguistic variants)
3. Unzip the images and you are ready to use the data in your local environment.