Prinasi/HDR-4D-Real

HDR-4D: Benchmark for Dynamic Novel View Synthesis in High Dynamic Range

This dataset repository contains the official benchmark datasets (HDR-4D-Syn and HDR-4D-Real) introduced in the ICLR 2026 paper:

Dynamic Novel View Synthesis in High Dynamic Range
Paper: arXiv:2410.02476
Official Codebase: Surrey-UP-Lab/HDR-4DGS

---

📌 Dataset Overview

High Dynamic Range Dynamic Novel View Synthesis (HDR DNVS) aims to reconstruct 4D dynamic radiance fields from sparse, multi-exposure Low Dynamic Range (LDR) observations while synthesizing photorealistic High Dynamic Range (HDR) views across continuous viewpoints and time steps.

This benchmark provides two complementary suites designed for both quantitative benchmarking and real-world evaluation:

Benchmark Subset	Scenes Count	Total Size	Modality	Primary Use
HDR-4D-Syn	8 scenes	~27 GB	Synthetic multi-exposure video + GT HDR frames	Controlled quantitative evaluation & benchmark
HDR-4D-Real	4 scenes	~14 GB	Multi-camera indoor capture (6x iPhone 14 Pro)	Real-world dynamic scene validation

---

📂 Dataset Hierarchy & Structure

Each scene is packaged as an independent .zip archive for fast, modular, and resumable downloads.

HDR-4DGS/
├── README.md
├── HDR-4D-Real/
│   ├── bed.zip             (~4.0 GB)
│   ├── excavator.zip       (~6.8 GB)
│   ├── tank.zip            (~2.4 GB)
│   └── toys.zip            (~1.1 GB)
└── HDR-4D-Syn/
    ├── airplane.zip        (~3.6 GB)
    ├── deer.zip            (~818 MB)
    ├── hook.zip            (~5.4 GB)
    ├── jump.zip            (~3.9 GB)
    ├── lego.zip            (~2.8 GB)
    ├── mutant.zip          (~4.7 GB)
    ├── standup.zip         (~4.8 GB)
    └── tank.zip            (~1.3 GB)

Extracted File Structure

1. HDR-4D-Syn Scenes

<scene_name>/
├── train/
│   ├── exr/                # High-fidelity Ground Truth HDR frames (.exr)
│   └── png/                # Multi-exposure LDR training frames (.png)
├── test/
│   ├── exr/                # Ground Truth evaluation HDR frames (.exr)
│   └── png/                # Ground Truth evaluation LDR frames (.png)
├── transforms_train.json   # Camera parameters & time-stamped poses for training
└── transforms_test.json    # Camera parameters & time-stamped poses for testing

2. HDR-4D-Real Scenes

<scene_name>/
├── exr/                    # Fused Ground Truth HDR frames (.exr)
├── images/                 # Synchronized multi-camera LDR captures (.jpeg/.png)
├── transforms_train.json   # Calibrated training camera poses & intrinsics
└── transforms_test.json    # Test camera trajectory & evaluation settings

---

🚀 Download Guide

1. Using huggingface-cli (Recommended)

Install huggingface_hub

pip install -U "huggingface_hub[cli]"

Download Entire Dataset

huggingface-cli download Prinasi/HDR-4DGS \
    --repo-type dataset \
    --local-dir ./HDR-4DGS

Download Specific Scene (Selective Download)

To download only a specific scene (e.g. HDR-4D-Syn/lego.zip or HDR-4D-Real/bed.zip):

# Download synthetic Lego scene
huggingface-cli download Prinasi/HDR-4DGS \
    HDR-4D-Syn/lego.zip \
    --repo-type dataset \
    --local-dir ./HDR-4DGS

# Download real-world Bed scene
huggingface-cli download Prinasi/HDR-4DGS \
    HDR-4D-Real/bed.zip \
    --repo-type dataset \
    --local-dir ./HDR-4DGS

(Optional for users in Mainland China): You can prepend HF_ENDPOINT=https://hf-mirror.com before the command for accelerated download.

---

2. Using Python API

from huggingface_hub import hf_hub_download, snapshot_download

# Download a single scene archive
file_path = hf_hub_download(
    repo_id="Prinasi/HDR-4DGS",
    filename="HDR-4D-Syn/lego.zip",
    repo_type="dataset",
    local_dir="./HDR-4DGS"
)

# Download the full dataset
snapshot_download(
    repo_id="Prinasi/HDR-4DGS",
    repo_type="dataset",
    local_dir="./HDR-4DGS"
)

---

3. Extracting Archives

Once downloaded, extract archives into your target workspace:

cd ./HDR-4DGS

# Extract all scenes in place
find . -name "*.zip" -execdir unzip -q {} \;

---

📷 Camera & Metadata Conventions

The camera parameters follow standard NeRF / Blender convention formatted in JSON:

• camera_angle_x: Horizontal field of view in radians.
• frames: List of dynamic frame metadata:
  • file_path: Relative path to the image/exr file.
  • transform_matrix: 4x4 camera-to-world transformation matrix.
  • time / index: Dynamic sequence temporal index.

---

📖 Citation

If you use this benchmark in your research, please cite the original ICLR 2026 paper:

@inproceedings{hdr4dgs2026,
  title     = {Dynamic Novel View Synthesis in High Dynamic Range},
  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.