XDen-1K (Preview Release)

This is a preview release of XDen-1K, containing 100 out of 1000 real-world objects.

XDen-1K is a large-scale, multi-modal dataset of real-world objects designed for physical property estimation, especially volumetric density fields reconstructed from X-ray observations

Each real object is captured with:

• category
• real-world scale
• image
• High-resolution 3D meshes with real scale
• Part-level annotations
• Biplanar X-ray scans
• Optimized volumetric density field, reconstructed from Bi-planar X-ray scans.

Data structure

dataset/
├── 01/
│   ├── ori_image.jpg            # RGB image
│   ├── mesh.glb                 # 3D mesh (real scale)
│   ├── density_pc.ply           # point cloud with per-point density
│   ├── optimized_lac.nii.gz      # optimized LAC volume
│   ├── lat.nii.gz               # lateral X-ray scan
│   ├── pa.nii.gz                # PA X-ray scan
│   └── part_lac_density.json     # per-part LAC and density values
├── 02/
│   └── ...
└── ...

The category and scale are in the metadata.json file

metadata.json

{
  "01": { 
    "category": "xx",            # Category the object belongs
    "scale": [0.0, 0.0, 0.0]     # scale of the object (cm)
  },
  "02": {
    "category": "yy",
    "scale": [0.0, 0.0, 0.0]
  }
  ...
}