HDR-NSFF: HDR-GoPro Dataset

Paper: HDR-NSFF: Neural Scene Flow Fields for Dynamic HDR Radiance Fields — ICLR 2026

Project Page: https://shin-dong-yeon.github.io/HDR-NSFF/ GitHub: https://github.com/kaist-ami/HDR-NSFF

Authors: Shin Dong-Yeon, Kim Jun-Seong, Kwon Byung-Ki, 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
• Held-out multi-view ground truth for novel-view synthesis evaluation (see gt/)

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_touch	Ball touching/catching
bear_thread	Thread interaction scene
big_jump	Large jumping motion

---

Data Structure

Each scene contains a dense/ directory with the training data and a gt/ directory with the held-out multi-view ground truth used for evaluation.

{scene}/
├── dense/                         # Training data (multi-exposure LDR + annotations)
│   ├── 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
│
└── gt/                            # Test / evaluation ground truth (held-out views)
    ├── mv_images/                 # Multi-view GT RGB frames (PNG)
    │   └── {frame}/               # Timestep index: 00000–00035 (36 frames)
    │       └── {view}.png         # Camera/view index: 00001–00009 (9 views)
    └── mv_masks/                  # Foreground motion masks, same layout as mv_images

The gt/ ground truth provides, for every scene, 36 timesteps × 9 multi-view captures (324 RGB frames and 324 corresponding motion masks per scene) at the original 1930 × 1081 resolution. Use these multi-view images to evaluate novel-view synthesis and HDR rendering quality (e.g. PSNR / SSIM / LPIPS), and the masks to compute foreground/dynamic-region metrics.

---

Usage

from huggingface_hub import hf_hub_download, snapshot_download

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

# Download only the evaluation ground truth for all scenes
snapshot_download(
    repo_id="SHlNDY/HDR-NSFF",
    repo_type="dataset",
    allow_patterns="*/gt/*",
    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:

@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) license.