S2R-HDR

🌐Homepage | 📖Arxiv | GitHub

✨Dataset Summary

S2R-HDR is a large-scale synthetic dataset for high dynamic range (HDR) reconstruction tasks. It contains 1,000 motion sequences, each comprising 24 images at 1920×1080 resolution, with a total of 24,000 images. To support flexible data augmentation, all images are stored in EXR format with linear HDR values. The dataset is rendered using Unreal Engine 5 and our custom pipeline built upon XRFeitoria, encompassing diverse dynamic elements, motion patterns, HDR-compatible scenes, and varied lighting conditions. Beyond the core imagery, we additionally provide per-frame rendered auxiliary data including optical flow, depth maps, surface normals, and diffuse albedo information, significantly expanding S2R-HDR's potential applications across various computer vision tasks.

Note: The other half of the data is stored in a separate repository: https://huggingface.co/datasets/iimmortall/S2R-HDR-2.

HDRs	Optical Flow	Depth	Diffuse	Normal

🔔News

• 🔥 We are pleased to announce the official release of our S2R-HDR dataset. We welcome the utilization of this resource and the provision of valuable feedback for continuous improvement.

🏭Dataset Structure

# S2R-HDR
./
├── scene_0_FM
│   ├── camera_params
│   ├── img
│   │   ├── 0000.exr
│   │   ├── [...]
│   │   ├── 0023.exr
│   ├── diffuse
│   │   ├── [...]
│   ├── flow
│   │   ├── [...]
│   ├── depth
│   │   ├── [...]
│   ├── normal
│   │   ├── [...]
├── scene_1_LM_shake
├── [...]

⚠ Note: In the suffix of the sequence name, LM/FM means that the motion mode of this sequence is Local Motion/Full Motion. Shake/No Shake means whether there is simulated camera shake in this sequence.

🏆Dataset Details

Dataset Construction

We implemented three targeted technical enhancements to enhance our synthetic dataset's photorealistic quality, content diversity, and scale.

• A custom UE5 rendering pipeline was designed to preserve radiometrically linear data by disabling built-in tone mapping and gamma correction operations throughout the rendering process.
• We incorporated Perlin noise-based camera shake simulation to replicate the vibrations and instabilities inherent in physical camera systems during real-world capture scenarios.
• A comprehensive asset library was developed containing: 1) animated foreground elements (animals, pedestrians, vehicles); 2) environment sets covering both interior and exterior spaces; 3) distinct HDR lighting configurations simulating daylight, twilight, and nighttime conditions with detailed quantitative breakdowns provided as follows.

✏️Citation

@article{wang2025s2r,
  title={S2R-HDR: A Large-Scale Rendered Dataset for HDR Fusion},
  author={Wang, Yujin and Wu, Jiarui and Bian, Yichen and Zhang, Fan and Xue, Tianfan},
  journal={arXiv preprint arXiv:2504.07667},
  year={2025}
}