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