ShapeNet

A large-scale library of 3D object meshes organized by WordNet categories, with semantic annotations.

ShapeNet was released in 2015 by researchers at Stanford, Princeton, and other institutions, containing more than 3 million 3D models, about 220,000 of which are sorted into 3,135 WordNet categories with annotations for canonical orientation, parts, symmetry planes, and real-world size. The most commonly used subset, ShapeNetCore, has about 51,300 models across 55 common categories; ShapeNetSem has fewer models but finer annotation. It originally served graphics and vision research, and later became a standard object source for robot grasping and simulation — populating a simulator with objects of varied shape at scale, or automatically generating grasp annotations. NVIDIA's ACRONYM grasp dataset uses ShapeNetSem meshes, annotating 17.7 million parallel-jaw grasps inside physics simulation. It is now available through Hugging Face and restricted to non-commercial research and educational use.

ExampleACRONYM takes cup and bowl meshes from ShapeNetSem and tests a large number of gripper grasp poses on each in simulation, using the successes and failures as training labels for a grasp network.

Also called

ShapeNetCore, ShapeNetSem

Related

Objaverse · ACRONYM: A Large-Scale Grasp Dataset Based on Simulation · Simulation Assets · YCB Object and Model Set · PartNet-Mobility · Domain Randomization

Sources

ShapeNet: An Information-Rich 3D Model Repository
ShapeNet/ShapeNetCore (Hugging Face)
ACRONYM: A Large-Scale Grasp Dataset Based on Simulation