ProcessedData
收藏资源简介:
UCBProject/ProcessedData数据集是Affordance2Grasp流水线第一阶段(Phase 1A)的输出结果,专为机器人抓取任务而设计。该数据集包含两个核心部分:training_fp/目录下的训练就绪数据,用于训练PointNet++接触预测网络(Phase 2);以及human_prior_fp/目录下的推理就绪数据,用于抓取采样、仿真执行和Sim2Real部署(Phase 3)。数据来源于两个公开数据集:OakInk(包含100个对象,对象ID如A01001、C*、O*等)和DexYCB(包含20个对象,对象ID如ycb_dex_01至ycb_dex_20)。每个对象的数据以HDF5格式存储,包含点云(4096个表面采样点,三维坐标,float32)、法线(每个点的单位法向量,float32)、人类先验接触概率(每个点的接触概率,经高斯平滑和逐对象最大归一化,float32)、机器人地面真值(全零占位符,float32)、力中心(接触概率≥80百分位的网格顶点质心,float32)以及对象ID属性(字符串)。数据集适用于机器人抓取中的接触预测、抓取规划等任务,原始数据基于DexYCB(NVIDIA,CC-BY-NC-4.0许可)和OakInk(CVPR 2022)重建的SAM3D对象网格生成。
The UCBProject/ProcessedData dataset is the output of Phase 1A, the first stage of the Affordance2Grasp pipeline, and is specifically designed for robotic grasping tasks. This dataset comprises two core parts: the training-ready data under the training_fp/ directory, which is used to train the PointNet++ contact prediction network (Phase 2); and the inference-ready data under the human_prior_fp/ directory, which is applied to grasp sampling, simulation execution, and Sim2Real deployment (Phase 3). The dataset is sourced from two public datasets: OakInk (containing 100 objects with object IDs such as A01001, C*, O*, etc.) and DexYCB (containing 20 objects with object IDs ranging from ycb_dex_01 to ycb_dex_20). Data for each object is stored in HDF5 format, including point clouds (4096 surface-sampled points with 3D coordinates, float32), normals (unit normal vectors for each point, float32), human prior contact probabilities (contact probability for each point, processed via Gaussian smoothing and per-object max normalization, float32), robot ground truth (all-zero placeholder, float32), force centers (centroids of grid vertices where contact probability ≥ 80th percentile, float32), and the object ID attribute (string). This dataset is applicable to tasks such as contact prediction and grasp planning in robotic grasping. The raw data is generated from SAM3D object meshes reconstructed based on DexYCB (NVIDIA, CC-BY-NC-4.0 license) and OakInk (CVPR 2022).
数据集概述:UCBProject / ProcessedData
数据集简介
该数据集是 Affordance2Grasp 管线的 Phase 1A 输出结果,供下游阶段使用:
- Phase 2(训练主方法 PointNet++ 接触预测网络)读取
training_fp/{dataset}/{object}.hdf5。 - Phase 3(推理、抓取采样、仿真执行、Sim2Real 部署)从
human_prior_fp/{object}.hdf5读取每个物体的先验 affordance 信息。
数据目录结构
training_fp/ oakink/ 100 个文件 (A* C* O* S* Y* — OakInk 物体 ID) dexycb/ 20 个文件 (ycb_dex_01 ... ycb_dex_20 — DexYCB 抓取物体) human_prior_fp/ A01001.hdf5 ... Y35037.hdf5 100 个 oakink 文件 ycb_dex_01.hdf5 ... ycb_dex_20.hdf5 20 个 dexycb 文件
HDF5 数据格式
training_fp/{ds}/{obj}.hdf5(训练就绪格式)
| 键 | 形状 | 数据类型 | 含义 |
|---|---|---|---|
point_cloud |
(4096, 3) | float32 | SAM3D 物体网格的 4096 个表面采样点(单位:米,网格规范坐标系) |
normals |
(4096, 3) | float32 | 每个点的单位法向量 |
human_prior |
(4096,) | float32 | 每个点的接触概率,范围 [0, 1](高斯平滑,每个物体最大归一化) |
robot_gt |
(4096,) | float32 | 全零占位符(该实验设置无机器人真实数据) |
force_center |
(3,) | float32 | 网格顶点中 contact_smooth >= 80th 百分位数 的质心 |
属性 object |
str | — | 物体 ID(如 ycb_dex_14 或 A01001) |
human_prior_fp/{obj}.hdf5(推理就绪格式,仅按物体索引)
与上述五个字段相同,内容与 training_fp/{ds}/{obj}.hdf5 一致,但跨数据集扁平化,以便 inference/predictor.py 仅通过物体 ID 读取。
数据来源与生成
- 生成流程:
data/batch_depth_pro.py→data/batch_haptic.py→tools/batch_obj_pose.py→data/batch_align_mano_fp.py - 原始数据源:
- DexYCB(NVIDIA,CC-BY-NC-4.0 许可)
- OakInk(CVPR 2022)
- Phase 1A 运行环境: 实验室 RTX 5090
- OakInk 完成于 2026-05-10
- DexYCB 完成于 2026-05-14
- 未包含的数据集: HO3D-v3 和 ARCTIC 暂未纳入,后续将在合作伙伴完成相应数据集的 Phase 1A 后添加。
上游数据
上游物体网格(SAM3D 重建)存储在相关仓库 UCBProject/ObjMesh 中。




