SyMTRS: Synthetic Multi-Task Remote Sensing Dataset

SyMTRS is a synthetic aerial imagery dataset generated using Unreal Engine 5 (UE5) to support multi-task learning in computer vision and generative AI research. The dataset is designed to benchmark and advance models across depth estimation, domain adaptation, and super-resolution — all within a consistent, high-quality simulated environment.

It also provides a valuable resource for generative AI applied to aerial and remote sensing imagery.

---

🌍 Overview

SyMTRS (Synthetic Multi-Task Remote Sensing) provides photorealistic aerial scenes rendered in UE5, enabling precise ground truth generation that is difficult or expensive to obtain in real-world aerial data.

The dataset is especially useful for:

• Multi-task learning research
• Synthetic-to-real domain transfer
• Training data for generative aerial imagery models
• Benchmarking perception models under controlled environmental changes

---

🧠 Supported Tasks

Domain Adaptation (Day → Night)

Each aerial scene includes paired daytime and nighttime renders, enabling research on visual domain shifts.

Use cases

• Robust perception under illumination changes
• Unsupervised domain adaptation
• Nighttime aerial monitoring systems

---

Super-Resolution

The dataset includes multiple resolution scales to support single-image super-resolution:

Scale	Description
×2	Moderate upscaling
×4	High upscaling
×8	Extreme upscaling

Use cases

• Enhancing low-resolution satellite or drone imagery
• Improving detail recovery in aerial scenes
• Training diffusion or GAN-based upscalers

---

Why Synthetic Aerial Data?

Real aerial datasets often lack:

• Accurate depth ground truth
• Perfectly aligned day/night pairs
• Multi-scale image consistency

SyMTRS solves these issues through simulation, providing:

• Pixel-perfect labels
• Controlled environmental variation
• Scalable data generation

---

Relevance to Generative AI

SyMTRS is not only a perception dataset — it is also well-suited for generative modeling in aerial imagery:

• Training diffusion or GAN models for aerial scene synthesis
• Learning structured scene representations
• Data augmentation for remote sensing
• Style and illumination transfer between domains

---

(SOON) Monocular Depth Estimation

High-quality depth maps are rendered directly from the UE5 simulation engine, providing accurate ground truth for training and evaluation.

Use cases

• 3D scene understanding
• Terrain reconstruction
• Urban structure modeling
• Navigation and mapping

---

📂 Dataset Structure

The dataset is organized by scene, task, and resolution level. A typical structure may look like:

SyMTRS/
│
├── hr/                  # ORIGINAL RAW IMAGES
│   ├── RS.0.png
│   └── ...
│
├── night/               # NIGHT VERSION OF IMAGES
│   ├── RS.0.png
│   └── ...
├── depth/               # DEPTH IMAGES
│   ├── RS.depth.0.npy
│   └── ...
├── lr/                  # BICUBIC DOWNSAMPLED IMAGES
│   ├── x2/
│   │   ├── RS.0.png
│   │   └── ...
│   ├── x4/
│   │   ├── RS.0.png
│   │   └── ...
│   ├── x8/
│   │   ├── RS.0.png
└   └   └── ...

---

🔬 Potential Research Directions

• Joint depth estimation + super-resolution models
• Domain-robust aerial perception systems
• Multi-task transformers for remote sensing
• Synthetic-to-real transfer learning
• Generative aerial world models

---

📊 Dataset Size

(SOON)

---

📜 License

This dataset is released under the Apache 2.0 License, allowing both academic and commercial use with proper attribution.

---

🤝 Citation

A PAPER WILL BE RELEASED.

---

🔗 Dataset Link

Hugging Face:
https://huggingface.co/datasets/safouaneelg/SyMTRS

---