Official dataset release for PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding.

• 2025-10-20: PartNeXt is on arXiv.

• Provide dataset toolkit.
• Provide example point cloud sampling code.
• Release dataset toolkit V2 supporting parse hierarchy semantics.
• Release benchmark code and data split.
• Release annotation platform code.
• Provide the index from PartNeXt to the original dataset.
• Provide template hierarchy and guidelines on using it.
• Release filtered data.

Our partnext dataset contains two parts:

• 3D meshes in .glb format, download from https://huggingface.co/datasets/AuWang/PartNeXt_mesh.
• 3D part and hierarchy annotations, download from https://huggingface.co/datasets/AuWang/PartNeXt.

You can download the dataset with huggingface-cli:

hf download --repo-type dataset AuWang/PartNeXt_mesh --local-dir /your/own/path
hf download --repo-type dataset AuWang/PartNeXt --local-dir /your/own/path

PartNeXt dataset toolkit supports loading meshes and annotations, and getting part geometries. This is an early version; we will support parse hierarchy semantics in the next version toolkit (hope to be done around Nov 2025).

We have uploaded our dataset toolkit to PyPI, so that you can directly install the toolkit by

pip install partnext

If you want to install the toolkit from source or refer to the code, you can clone the toolkit repo

git clone https://github.com/AuthorityWang/PartNeXt_lib.git
cd PartNeXt_lib
pip install -e .

Please refer to example/toolkit_example.py.

We propose 2 benchmarks for the PartNeXt dataset:

• Class-Agnostic 3D Part Segmentation.
• Part-Centric 3D Question Answering.

The data split and evaluation code will be provided soon. Please contact us if you need to get the code now.

Currently, our dataset toolkit is very simple. If you want to parse the part semantic hierarchy and get non-leaf node parts, you can refer to the format and try to load the data with your own code.

We store our 3D meshes in .glb format, folder structure follows Objaverse

PartNeXt_mesh/
├── glbs
│   ├── 000-000
│   │   ├── 000074a334c541878360457c672b6c2e.glb
│   │   ├── 0000ecca9a234cae994be239f6fec552.glb
│   │   └── ...
│   ├── 000-001
│   ├── ...

We store our data in arrow format, all info stored as a string. After loading the dataset with HuggingFace's datasets library, the data has the following columns:

• model_id: Model UUID, same as glb name. (Object from Objaverse uses the original Objaverse ID)
• type_id: Subfolder name, follow the objaverse structure.
• user_id: Annotator ID.
• anno_time: Time consumed for annotation, in seconds.
• mesh_face_num: As each object can consist of multiple meshes, we store the number of faces for each mesh.
• masks: Leaf node's part masks, corresponding to the ​​finest granularity​​ parts.
• hierarchyList: Hierarchy list, each node can have children and semantics, is a tree structure like PartNet.

For mesh_face_num masks hierarchyList, we give a further explanation on data structure and show examples.

{
    "0": 2416,
    "1": 672,
    "2": 2
}

{
    "0": {
        "0": [0, 1, 2, 3, 4, ...], 
        "1": [221, 222, 223, ...]
    }, 
    "1": {
        "0": [5, 6, 7, 8, 9, ...], 
        "1": [220, 221, 222, ...]
    }, 
    ...
}

[
  {
    "name": "Table",
    "nodeId": 0,
    "refNodeId": 0,
    "children": [
      {
        "name": "Standard Table",
        "nodeId": 1,
        "refNodeId": 1,
        "children": [
          {
            "name": "Tabletop",
            "nodeId": 2,
            "refNodeId": 2,
            "children": [
              {
                "name": "Surface Panel",
                "nodeId": 3,
                "refNodeId": 3,
                "maskId": 0
              }
            ]
          },
          ...
        ]
      },
      ... 
    ]   
  }
]

We give some examples under the example folder using our PartNeXt dataset. You can refer to this code to better understand the dataset.

Point-SAM needs part part-level point cloud to train a promptable 3D part segmentation model. We provide a sample code to sample part point clouds from the PartNeXt dataset, and save them in Point-SAM's format.

This example requires raw annotation in JSON format, which can be downloaded from https://huggingface.co/datasets/AuWang/PartNeXt_raw.

You can open an issue to ask for more examples on specific tasks.

Though we perform a check on the annotation, we admit that PartNeXt still has some errors in the annotation. We will filter some error annotations in the future. If you found error in the annotation, please open an issue or contact us.

Our PartNeXt dataset is based on Objaverse, ABO, 3D-Future, thanks for these awesome datasets. If there is any license issue, please contact us, and we will remove the data.

Thanks to Benyuan AI data for data annotation.

If you find our dataset useful in your research, please consider citing our paper.

BibTex Coming Soon