• t2_345327
  • 1.5MB
  • zip
  • 0
  • VIP专享
  • 0
  • 2022-06-14 03:41
KeypointNet 完整的数据集现在可用! KeypointNet是一个大规模且多样化的3D关键点数据集,通过利用基于ShapeNet模型的众多人工注释,包含来自16个对象类别的83,231个关键点和8,329个3D模型。 我们的论文可在上获得并被CVPR 2020接受。您可以在以下位置浏览我们的数据集 。 消息! 现在发布了两个有趣的无监督关键点检测器: (无序但SE(3)不变)和 (非SE(3)不变但有序)! 变更记录 有关更新的数据集信息,请参阅 关键点数据 数据集可以从或下载。 带注释的JSON数据放置在注释下。 此外,我们还提供了下PCD的每个ShapeNet模型采样点云(2048点)。 我们已经为飞机(1022型号),浴缸(492型号),床(146型号),瓶子(380型号),瓶盖(38型号),汽车(1002型号),椅子(999型号),吉他等标签进行了处理和清洗。
# KeypointNet **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 and is accepted to CVPR 2020. **You can explore our dataset on <a href="" target="_blank" rel='nofollow' onclick='return false;'></a>**. **News! Two interesting unsupervised keypoint detectors: [UKPGAN]( (unordered but SE(3) invariant) and [SkeletonMerger]( (not SE(3) invariant but ordered) are now released!** # Change Logs For updated dataset information, see [Change Log]( # Keypoint Data Dataset can be downloaded from <a href="" target="_blank" rel='nofollow' onclick='return false;'>Google Drive</a> or <a href="!Aj0NuSsDz6hDyF3LT3xaPkXK9DXC?e=kcrfSg" target="_blank" rel='nofollow' onclick='return false;'>OneDrive</a>. 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 ****) 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. <!-- ![pcd](examples/captures/pcd.png){:height="360px" width="160px"} ![obj](examples/captures/obj.png){:height="360px" width="160px"} ![ply](examples/captures/ply.png){:height="360px" width="160px"} --> ## Data format ```javascript [ ..., { "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 } }, ... ] } }, ... ] ``` # Examples Example scripts on reading and visualizing keypoints on both point clouds and triangle meshes are placed under **examples**. # Keypoint Detection Tasks Keypoint saliency and correspondence training and evaluation baselines for various backbones are placed under **benchmark_scripts**. For more details, please refer to **benchmark_scripts/**. # Data Splits train/val/test splits are placed under **splits**. Each line is formatted as `[class_id]-[model_id]`. # Citation 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} } ``` # TODOs - [x] clean labels for more classes - [x] add colored pcds/meshes - [x] a browser interface to explore dataset