RealX3D: A Physically-Degraded 3D Benchmark for Multi-view Visual Restoration and Reconstruction

RealX3D is a real-world benchmark dataset for multi-view 3D reconstruction under challenging capture conditions. It provides multi-view RGB images (both processed JPEG and Sony RAW), COLMAP sparse reconstructions, and high-precision 3D ground-truth geometry (point clouds, meshes, and rendered depth maps) across a diverse set of scenes and degradation types.

🌙 Low Light	💨 Smoke

✨ Key Features

• 9 real-world degradation conditions: defocus (mild/strong), motion blur (mild/strong), low light, smoke, reflection, dynamic objects, and varying exposure.
• Full-resolution (~7000×4700) and quarter-resolution (~1800×1200) JPEG images with COLMAP reconstructions.
• Sony RAW (ARW) sensor data with complete EXIF metadata for 7 conditions.
• Per-frame metric depth maps rendered from laser-scanned meshes.
• Camera poses and intrinsics in both COLMAP binary format and NeRF-compatible transforms.json.

📁 Dataset Structure

RealX3D/
├── data/              # Full-resolution JPEG images + COLMAP reconstructions
├── data_4/            # Quarter-resolution JPEG images + COLMAP reconstructions
├── baseline_results/  # Baseline methods rendering results on data_4 for direct download
├── data_arw/          # Sony RAW (ARW) sensor data
├── pointclouds/       # 3D point clouds, meshes, and metric depth maps
└── scripts/           # Utilities scripts

🚀 Release Status

• data/ — Full-resolution JPEG images + COLMAP
• data_4/ — Quarter-resolution JPEG images + COLMAP
• baseline_results/ - Baseline rendering results
• data_arw/ — Sony RAW (ARW) sensor data
• pointclouds/ — 3D ground-truth geometry (point clouds, meshes, depth maps)

🌧️ Capture Conditions

Condition	Description
defocus_mild	Mild defocus blur
defocus_strong	Strong defocus blur
motion_mild	Mild motion blur
motion_strong	Strong motion blur
dynamic	Dynamic objects in the scene
reflection	Specular reflections
lowlight	Low-light environment
smoke	Smoke / particulate occlusion
varyexp	Varying exposure

🏛️ Scenes

Akikaze, BlueHawaii, Chocolate, Cupcake, GearWorks, Hinoki, Koharu, Laboratory, Limon, MilkCookie, Natsume, Popcorn, Sculpture, Shirohana, Ujikintoki

---

📸 data/ — Full-Resolution JPEG Images

Full-resolution JPEG images and corresponding COLMAP sparse reconstructions, organized by condition → scene.

Per-Scene Directory Layout

data/{condition}/{scene}/
├── train/                    # Training images (~23–31 frames)
│   ├── 0001.JPG
│   └── ...
├── val/                      # Validation images (~23–31 frames)
│   └── ...
├── test/                     # Test images (~4–6 frames)
│   └── ...
├── transforms_train.json     # Camera parameters & poses (training split)
├── transforms_val.json       # Camera parameters & poses (validation split)
├── transforms_test.json      # Camera parameters & poses (test split)
├── point3d.ply               # COLMAP sparse 3D point cloud
├── colmap2world.txt          # 4×4 COLMAP-to-world coordinate transform
├── sparse/0/                 # COLMAP sparse reconstruction
│   ├── cameras.bin / cameras.txt
│   ├── images.bin / images.txt
│   └── points3D.bin / points3D.txt
├── distorted/sparse/0/       # Pre-undistortion COLMAP reconstruction
└── stereo/                   # MVS configuration files

📐 transforms.json Format

Each transforms_*.json file contains shared camera intrinsics and per-frame extrinsics following Blender Dataset format, for example:

{
  "camera_angle_x": 1.295,
  "camera_angle_y": 0.899,
  "fl_x": 4778.31,
  "fl_y": 4928.04,
  "cx": 3649.23,
  "cy": 2343.41,
  "w": 7229.0,
  "h": 4754.0,
  "k1": 0, "k2": 0, "k3": 0, "k4": 0,
  "p1": 0, "p2": 0,
  "is_fisheye": false,
  "aabb_scale": 2,
  "frames": [
    {
      "file_path": "train/0001.JPG",
      "sharpness": 25.72,
      "transform_matrix": [[...], [...], [...], [...]]
    }
  ]
}

All distortion coefficients are zero (images are pre-undistorted).

🖼️ Image Specifications

• Format: JPEG
• Resolution: ~7000 × 4700 pixels (varies slightly across scenes)
• Camera: Sony ILCE-7M4 (α7 IV)
• Camera Model: PINHOLE (pre-undistorted)

---

📸 data_4/ — Quarter-Resolution JPEG Images (Used for 2026 NTIRE-3DRR Challenge)

Identical directory structure to data/, with images downsampled to 1/4 resolution (~1800 × 1200 pixels). Camera intrinsics (fl_x, fl_y, cx, cy, w, h) in the transforms.json files are adjusted accordingly. All 9 capture conditions and their scenes are included.

---

📷 data_arw/ — Sony RAW Data

Sony ARW (TIFF-wrapped RAW) sensor data preserving full EXIF metadata.

Differences from data/

• Image format: .ARW (~33–35 MB per frame)
• 7 conditions available: defocus_mild, defocus_strong, dynamic, lowlight, reflection, smoke, varyexp (motion blur conditions are excluded)

Per-Scene Directory Layout

data_arw/{condition}/{scene}/
├── train/              # ARW raw images
├── val/
├── test/
└── sparse/0/           # COLMAP sparse reconstruction

---

📍 pointclouds/ — 3D Ground Truth

High-precision 3D geometry ground truth, organized directly by scene name (geometry is shared across capture conditions for the same scene).

Per-Scene Directory Layout

pointclouds/{scene}/
├── cull_pointcloud.ply   # Culled point cloud (view-frustum trimmed)
├── cull_mesh.ply         # Culled triangle mesh
├── colmap2world.npy      # 4×4 COLMAP-to-world transform (NumPy format)
└── depth/                # 16-bit Depth maps rendered from the mesh
    ├── 0001.png
    ├── 0002.png
    └── ...

The colmap2world.npy matrix aligns COLMAP reconstructions to the world coordinate system of the ground-truth geometry. The same transform is also stored as colmap2world.txt in the corresponding data/ directories.

---

📜 Citation

@article{liu2025realx3d,
  title   = {RealX3D: A Physically-Degraded 3D Benchmark for Multi-view
             Visual Restoration and Reconstruction},
  author  = {Liu, Shuhong and Bao, Chenyu and Cui, Ziteng and Liu, Yun
             and Chu, Xuangeng and Gu, Lin and Conde, Marcos V and
             Umagami, Ryo and Hashimoto, Tomohiro and Hu, Zijian and others},
  journal = {arXiv preprint arXiv:2512.23437},
  year    = {2025}
}

---

📄 License

This dataset is released under the MIT License.