Full dataset is available now! KeypointNet is a large-scale and diverse 3D keypoint dataset that contains 83,231 keypoints and 8,329 3D models from 16 object categories, by leveraging numerous human annotations, based on ShapeNet models. Our paper is available on https://arxiv.org/pdf/2002.12687.pdf and is accepted to CVPR 2020. You can explore our dataset on http://qq456cvb.github.io/keypointnet/explore/.

News! Two interesting unsupervised keypoint detectors: UKPGAN (unordered but SE(3) invariant) and SkeletonMerger (not SE(3) invariant but ordered) are now released!

For updated dataset information, see Change Log

Dataset can be downloaded from Google Drive or OneDrive. Annotated JSON data is placed under annotations. In addition, we provide sampled point clouds (2048 points) for each ShapeNet model under pcds.

We have processed and cleaned labels for airplane (1022 models), bathtub (492 models), bed (146 models), bottle (380 models), cap (38 models), car (1002 models), chair (999 models), guitar (697 models), helmet (90 models), knife (270 models), laptop (439 models), motorcycle (298 models), mug (198 models), skateboard (141 models), table (1124 models) and vessel (910 models).

UPDATE: we have managed to add color information onto sampled point clouds and keypoints. In addition, since processing raw ShapeNet obj file as colored triangle meshes is painful, we have generated corresponding ply files (named ShapeNetCore.v2.ply.zip) with vertex colors (diffuse texture color), for those are interested in dealing with triangle meshes. We believe color is an important source when learning from 3D geometries.

[
    ...,
    {  
        "class_id": "03001627",  // WordNet id
        "model_id": "88382b877be91b2a572f8e1c1caad99e",  // model id
        "keypoints": [
            {
                "xyz": [0.16, 0.1, 0.1],  // xyz coordinate of keypoint
                "rgb": [255, 255, 255], // rgb color of keypoint, uint8
                "semantic_id": 0,  // id of semantic meaning
                "pcd_info": {
                    "point_index": 0  // keypoint index on corresponding point cloud
                },
                "mesh_info": {  // mesh information for both obj and ply files
                    "face_index": 0,  // index of mesh face where keypoint lies
                    "face_uv": [0.2, 0.4, 0.4]  // barycentric coordinate on corresponding mesh face
                }
            },
            ...
        ],
        "symmetries": { // information of keypoint symmetries
            "reflection": 
            [
                {
                    "kp_indexes": [0, 1]  // keypoint indexes of a reflection symmetric group
                },
                ...
            ],
            "rotation":
            [
                {
                    "kp_indexes": [0, 1, 2, 3],  // keypoint indexes of a rotation symmetric group
                    "is_circle": true,  // true if this rotation symmtric group is a rounding circle
                    "circle": {
                        "center": [0.2, 0.5, 0.2],  // circle center
                        "radius": 0.32,  // circle radius
                        "normal": [0, 1.0, 0],  // normal of circle plane
                    }
                },
                ...
            ]
        }

    },
    ...
]

Example scripts on reading and visualizing keypoints on both point clouds and triangle meshes are placed under examples.

Keypoint saliency and correspondence training and evaluation baselines for various backbones are placed under benchmark_scripts. For more details, please refer to benchmark_scripts/README.md.

train/val/test splits are placed under splits. Each line is formatted as [class_id]-[model_id].

If you find KeypointNet data or code useful in your research, please consider citing:

@article{you2020keypointnet,
  title={KeypointNet: A Large-scale 3D Keypoint Dataset Aggregated from Numerous Human Annotations},
  author={You, Yang and Lou, Yujing and Li, Chengkun and Cheng, Zhoujun and Li, Liangwei and Ma, Lizhuang and Lu, Cewu and Wang, Weiming},
  journal={arXiv preprint arXiv:2002.12687},
  year={2020}
}

• clean labels for more classes
• add colored pcds/meshes
• a browser interface to explore dataset

Our KeypointNet is released under MIT license.