---
license: mit
task_categories:
- depth-estimation
- image-to-image
tags:
- space
- lunar
- simulation
- computer-vision
- robotics
- benchmark
- depth
size_categories:
- 100K<n<1M
language:
- en
---

# LunarSim: A Synthetic Dataset of the LuMon Benchmark

## 1. Dataset Summary
The **LunarSim** dataset is a synthetic collection of paired stereo imagery and pixel-aligned depth maps, explicitly designed for evaluating Monocular Depth Estimation (MDE) on lunar surfaces. 

It is a dataset introduced by our work **LuMon**, which was introduced to address the severe domain gap caused by harsh lunar shadows, textureless regolith, and the absence of atmospheric scattering. Captured within a high-fidelity Unity-based simulator, LunarSim isolates these extreme optical conditions to rigorously test the structural consistency and sim-to-real transfer capabilities of both metric and relative deep-learning architectures.

## 2. Dataset Structure & Details
The dataset contains approximately 3,228 synchronized image pairs (totaling ~1.9 GB), extracted from continuous simulation trajectories.

### Directory Structure
* **`left_camera/`**: Contains the input images captured from the rover's left optical camera.
* **`depth_image/`**: Contains the corresponding ground-truth depth maps.

### Technical Specifications
* **Resolution:** All images (both RGB and Depth) are `1280x720` pixels.
* **Synchronization:** Input images and their corresponding depth maps are perfectly paired using matching UNIX timestamps in their filenames (e.g., `left_1736612498.png` corresponds directly to `depth_1736612498.png`).
* **Depth Format (Important):** The depth maps are saved as **8-bit Grayscale PNGs** (`mode="L"` in PIL/OpenCV). Pixel values are integers ranging from `0` to `255`. Note that because this specific simulation split lacks an absolute metric scale, these 8-bit values represent **relative (normalized) depth** rather than absolute metric distances in meters.

## Citation
This dataset is a subset of the **LuMon Benchmark** suite, accepted at the **CVPR 2026 AI4Space Workshop**. If you use this dataset in your research, please cite:

```bibtex
@inproceedings{sekmen2026lumon,
  title={LuMon: A Comprehensive Benchmark and Development Suite with Novel Datasets for Lunar Monocular Depth Estimation},
  author={Aytac Sekmen and Fatih Gunes and Furkan Horoz and Umut Isik and Alp Ozaydin and Altay Topaloglu and Umutcan Ustundas and Alp Yeni and Ersin Soken and Erol Sahin and Gokberk Cinbis and Sinan Kalkan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year={2026}
}