SHlNDY/HDR-NSFF

---
license: cc-by-4.0
task_categories:
- video-to-video
- other
tags:
- HDR
- HDR-video
- dynamic-scene
- 4D-reconstruction
- novel-view-synthesis
- nerf
- scene-flow
- gopro
- multi-exposure
- ICLR2026
pretty_name: HDR-NSFF (HDR-GoPro Dataset)
size_categories:
- 1G<n<10G
---

# HDR-NSFF: HDR-GoPro Dataset

**Paper:** [HDR-NSFF: Neural Scene Flow Fields for Dynamic HDR Radiance Fields](https://arxiv.org/abs/2603.08313) — *ICLR 2026*

**Project Page:** [https://shin-dong-yeon.github.io/HDR-NSFF/](https://shin-dong-yeon.github.io/HDR-NSFF/)
**GitHub:** [https://github.com/kaist-ami/HDR-NSFF](https://github.com/kaist-ami/HDR-NSFF)

**Authors:** Dong-Yeon Shin, Jun-Seong Kim, Byung-Ki Kwon, Tae-Hyun Oh

---

## Abstract

We present HDR-NSFF, a method for reconstructing dynamic 4D scenes with HDR video rendering from multi-exposure monocular video. Our approach extends Neural Scene Flow Fields (NSFF) to jointly learn camera response functions (CRF), scene geometry, and temporal dynamics from bracketed exposure sequences captured by a GoPro camera. The reconstructed radiance field supports novel-view synthesis, bullet-time rendering, and HDR tone-mapping with physically accurate scene flow.

---

## Dataset Description

The **HDR-GoPro** dataset consists of dynamic outdoor and indoor scenes captured with a GoPro camera using automatic exposure bracketing. Each scene provides multi-exposure frames enabling HDR reconstruction.

- **12 scenes** of dynamic human activities
- **9 cameras** / exposure levels per scene (3-exposure bracketing × 3 positions)
- Multi-exposure LDR frames for HDR fusion
- Camera poses estimated via COLMAP
- Metric depth from Depth-Anything-V2
- Semantic optical flow from DINO-tracker
- Motion masks from SAM2

### Scenes

| Scene | Description |
|-------|-------------|
| `tumbler` | Person shaking a tumbler |
| `dog` | Dog running |
| `jumping_jack` | Jumping jacks exercise |
| `pointing_walk` | Person walking and pointing |
| `side_walk` | Side-view walking |
| `tube_toss` | Tossing a tube |
| `fire_extinguisher` | Fire extinguisher action |
| `laptop` | Laptop interaction |
| `bag` | Bag swinging |
| `ball` | Ball throwing/catching |
| `bear_thread` | Thread interaction scene |
| `big_jump` | Large jumping motion |

---

## Data Structure

```
{scene}/
└── dense/
├── images/ # Original LDR frames (JPEG)
├── images_{W}x{H}/ # Resized frames for training
├── motion_masks/ # Foreground motion masks (SAM2)
├── depth-anything/ # Metric depth maps (Depth-Anything-V2)
├── semantic_flow_i1/ # Per-frame-pair semantic flow (.npz)
├── dino-tracker/
│ └── semantic_flow/ # Raw DINO-tracker flow arrays (.npy)
└── poses_bounds.npy # LLFF-format camera poses & bounds
```

---

## Usage

```python
from huggingface_hub import hf_hub_download, snapshot_download

# Download a single scene
snapshot_download(
repo_id="SHlNDY/HDR-NSFF",
repo_type="dataset",
allow_patterns="tumbler/*",
local_dir="./data/hdr-gopro",
)

# Download only camera poses for all scenes
from huggingface_hub import HfFileSystem
fs = HfFileSystem()
pose_files = fs.glob("datasets/SHlNDY/HDR-NSFF/*/dense/poses_bounds.npy")
```

---

## Citation

If you use this dataset in your research, please cite:

```bibtex
@inproceedings{shin2026hdrnsff,
title = {HDR-NSFF: Neural Scene Flow Fields for Dynamic HDR Radiance Fields},
author = {Shin, Dong-Yeon and Kim, Jun-Seong and Kwon, Byung-Ki and Oh, Tae-Hyun},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2026},
}
```

---

## License

This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license.