LAE-1M

# LAE-1M: Locate Anything on Earth Dataset <p align="center"> <img src="https://jianchengpan.space/projects/LAE/assets/LAE-1M.png" alt="LAE-1M" width="600"/> </p> **LAE-1M** (Locate Anything on Earth - 1 Million) is a large-scale **open-vocabulary remote sensing object detection dataset** introduced in the paper *"Locate Anything on Earth: Advancing Open-Vocabulary Object Detection for Remote Sensing Community"* (AAAI 2025). It contains over **1M images** with **coarse-grained (LAE-COD)** and **fine-grained (LAE-FOD)** annotations, unified in **COCO format**, enabling **zero-shot** and **few-shot** detection in remote sensing. --- ## Dataset Details ### Dataset Description - **Curated by:** Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiahao Li, Danda Pani Paudel, Luc Van Gool, Xiaomeng Huang - **Funded by:** ETH Zürich, INSAIT (partial computing support) - **Shared by:** LAE-DINO Project Team - **Language(s):** Not language-specific; visual dataset - **License:** MIT License ### Dataset Sources - **Repository:** [GitHub - LAE-DINO](https://github.com/jaychempan/LAE-DINO) - **Paper:** [ArXiv 2408.09110](https://arxiv.org/abs/2408.09110), [AAAI 2025](https://ojs.aaai.org/index.php/AAAI/article/view/32672) - **Project Page:** [LAE Website](https://jianchengpan.space/LAE-website/index.html) - **Dataset Download:** [HuggingFace](https://huggingface.co/datasets/jaychempan/LAE-1M) --- ## Dataset Structure | Subset | # Images | # Classes | Format | Description | |-------------|-------------|-----------|-------------|----------------------------------------------| | LAE-COD | 400k+ | 20+ | COCO | Coarse-grained detection (AID, EMS, SLM) | | LAE-FOD | 600k+ | 50+ | COCO | Fine-grained detection (DIOR, DOTAv2, FAIR1M) | | LAE-80C | 20k (val) | 80 | COCO | Benchmark with 80 semantically distinct classes | All annotations are in **COCO JSON** format with bounding boxes and categories. --- ## Uses ### Direct Use - Open-Vocabulary Object Detection in Remote Sensing - Benchmarking zero-shot and few-shot detection models - Pretraining large vision-language models ### Out-of-Scope Use - Any tasks requiring personal or sensitive information - Real-time inference on satellite streams without further optimization --- ## Quick Start ```python from datasets import load_dataset # Load the dataset dataset = load_dataset("jaychempan/LAE-1M", split="train") # Access one example example = dataset[0] print(example.keys()) # image, annotations, category_id, etc. # Show the image (requires Pillow) from PIL import Image import io img = Image.open(io.BytesIO(example["image"])) img.show()