SyMTRS: Synthetic Multi-Task Remote Sensing Dataset

SyMTRS is a high-fidelity simulation dataset generated with Unreal Engine 5 (UE5), built on top of the MatrixCity project. It provides multi-modal, perfectly aligned data for aerial imagery research, including:

• High-resolution RGB images (2048×2048)
• Multi-scale low-resolution images
• Day/Night domain pairs
• Ground-truth depth maps

Project Page

🔗 Links

• Dataset: https://huggingface.co/datasets/safouaneelg/SyMTRS
• Code: https://github.com/safouaneelg/SyMTRS
• Paper https://arxiv.org/abs/2604.21801

---

🌍 Overview

SyMTRS enables multi-task learning for:

• Super-resolution
• Monocular depth estimation
• Domain adaptation (day ↔ night)
• Image-to-image translation
• Generative aerial modeling

Unlike real aerial datasets, SyMTRS provides:

• Pixel-perfect alignment across modalities
• Accurate depth ground truth
• Controlled lighting and environmental variation

---

🧠 Supported Tasks

1. Super-Resolution

The dataset provides bicubic downsampled images at multiple scales:

Scale	LR	HR
×2	1024×1024	2048×2048
×4	512×512	2048×2048
×8	256×256	2048×2048

---

2. Image-to-Image Translation / Domain Adaptation

Each scene includes:

• Daytime image (hr/)
• Nighttime image (night/)

Supports:

• CycleGAN (unpaired)
• pix2pix (paired)

---

3. Monocular Depth Estimation

Depth maps are stored as .npy files:

• Directly rendered from UE5
• Perfectly aligned with RGB images

---

📂 Dataset Structure

SyMTRS/
├── hr/
├── night/
├── depth/
└── lr/
├── x2/
├── x4/
└── x8/

Matching rules:

• RS.{id}.png ↔ RS.{id}.png
• RS.depth.{id}.npy ↔ RS.{id}.png

---

📦 Loading the Dataset

Using 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("safouaneelg/SyMTRS")
print(ds)

---

🔧 Official Dataloaders (Recommended)

Clone the repo:

git clone https://github.com/safouaneelg/SyMTRS
cd SyMTRS

Install dependencies:

pip install torch torchvision numpy pillow matplotlib tqdm timm datasets

---

Super-Resolution

from dataloader import SuperResDataset

dataset = SuperResDataset(
    lr_root="/path/to/SyMTRS/lr/x4",
    hr_root="/path/to/SyMTRS/hr",
)

---

Domain Adaptation

from dataloader import DomainAdaptDataset

dataset = DomainAdaptDataset(
    day_root="/path/to/SyMTRS/hr",
    night_root="/path/to/SyMTRS/night",
    paired=True,
)

---

Depth Estimation

from dataloader import DepthDataset

dataset = DepthDataset(
    depth_root="/path/to/SyMTRS/depth",
    hr_root="/path/to/SyMTRS/hr",
)

---

🚀 Training Code

Entry point:

python train_superresolution.py

Supported models:

• SRCNN
• Autoencoder
• SRGAN
• SwinIR

Example:

python train_superresolution.py \
  --lr_root /path/to/SyMTRS/lr/x4 \
  --hr_root /path/to/SyMTRS/hr \
  --model srcnn \
  --scale 4

---

🧪 Research Directions

• Multi-task learning (SR + depth)
• Synthetic-to-real transfer
• Domain-robust aerial perception
• Generative aerial world models
• Remote sensing data augmentation

---

⚠️ Limitations

• Synthetic dataset → domain gap with real-world imagery
• Requires adaptation for real satellite/drone deployment

---

📜 License

Apache 2.0

---

📖 Citation

@misc{elghazoualisymtrs,
      title={SyMTRS: Benchmark Multi-Task Synthetic Dataset for Depth, Domain Adaptation and Super-Resolution in Aerial Imagery}, 
      author={Safouane El Ghazouali and Nicola Venturi and Michael Rueegsegger and Umberto Michelucci},
      year={2026},
      eprint={2604.21801},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.21801}, 
}

---