mburgjr/GelSight-YoungsModulus

GelSight Young's Modulus Dataset ==============

Dataset of tactile images collected over grasping common objects labelled with the objects' Young's Moduli. All images are collected using GelSight Wedges, both with and without displacement markers. Associated code is available on GitHub.

Pre-print is available on arXiv.

Dataset Format ---------------

Each object is grasped a number of times, and each of these grasps is downsampled to 3 frames and shifted to create different data augmentations of the same grasp. Collected data is split into folders according to the following structure...

gelsight_youngs_modulus_dataset
└── {object_name}
    ├── metadata.json
    └── grasp={grasp_number}
        └── augmentation={augmentation_number}
             ├── RGB.pkl
             ├── depth.pkl
             ├── RGB_markers.pkl
             ├── depth_markers.pkl
             ├── forces.pkl
             ├── widths.pkl
             ├── elastic_estimate.pkl
             └── hertz_estimate.pkl

In the OBJECT_NAME folder, metadata for each object is provided in a .json file, including the object's shape, material, Young's modulus, and Shore hardness.

metadata = {
    'object_name':         {object_name}, # [str]
    'youngs_modulus':      0.0, # [Pa]
    'material':            '', # [str]
    'shape':               '', # [str]
    'shore_00_hardness':   None, # [/]
    'shore_A_hardness':    None, # [/]
    'shore_D_hardness':    None, # [/]
    'used_in_training':    True # [bool]
}

In the AUGMENTATION_NUMBER folder, data is provided in .pkl files. You will find the following named conventions...

• —Tactile RGB images (3 images, without markers): RGB.pkl
• —Tactile depth images (3 images, without markers): depth.pkl
• —Tactile RGB images (3 images, with markers): RGB_markers.pkl
• —Tactile depth images (3 images, with markers): depth_markers.pkl
• —Sampled grasping contact forces: forces.pkl
• —Sampled gripper widths: widths.pkl
• —Elastic analytical model estimate: elastic_estimate.pkl
• —Hertzian analytical model estimate: hertz_estimate.pkl

Dataset Diversity: ---------------

Shapes:

• —Cylinder: 20.2%
• —Rectangular: 15.3%
• —Sphere: 11.5%
• —Hexagonal: 2.1%
• —Irregular: 50.9%

Materials:

• —Rubber: 28.6%
• —Metal: 25.1%
• —Plastic: 24.0%
• —Wood: 8.4%
• —Foam: 6.3%
• —Other: 7.6%