Spherical-Data-Generation-For-3D-Meshes

Spherical Representation of Grasps and ShapeNet Datasets

Using ShapeNet and the Acroynm Datasets (depicted below), we create a spherical representation of the data through hemispherical radial ray-casting

Example ShapeNet + Acroynm baseline representation

See below an example of sphere enclosing object. Sphere displays ray-casting origin points. Note that only a hemisphere is used during data creation

Resulting Data:

Resulting Data is a spherical projection of the 3-D meshes.

The Data is organized as follows:

Channel 1: Spherical Depth Data, obtain through simple ray casting Image below shows the recording depth/hit-distance between the enclosing hemisphere and the object in question

Channel 2: Absolute Grasp Position Data, Mapped to nearest ray Image below shows the absolute grasp positional data, where points correspond to the nearest ray that a grasp would be located with.

Note that for the following orientation data. The values at each plotted position are based on the spherical coordinate system. To identify plane of rotation, grasp vector and orgination ray vector are used

Channel 3: Spherical Theta Value --> Used to obtain orientation of grasp

Channel 4: Spherical Phi Value --> Used to obtain orientation of grasp

Channel 5: Rotational Gamma Value --> Used to obtain grasp rotation

Gaussian Mixture Models (n_components) representations of data:

A primary objective of this dataset is to create a probability mapping of the position and orientation data. To do so, Gaussian mixture models were implemented.

Dense Raw Data

GMM Dense Data Positional

GMM Dense Data 3D

Sparse Data Positional

GMM Sparse Data Positional