Vision-Based Grasping Benchmark

The dataset

The 18 objects used in the database are displayed in the figure at the right. The objects are part of the KITObjectModels WebDatabase. 3Dmodels of all objects are available for the grasp simulation.

The objects have various shapes, sizes, colors, and textures. We recorded scenes with one object and with two objects. In the single-object case, we recorded the 18 different objects in eight different poses, four where the object stands upright, and four where the object lies down. In the double-object scenes, we have 9 combinations of objects, where the objects are in eight different configurations, four where the objects are placed apart, and four where the objects touch each other. All scenes are recorded in two conditions, placed on a non-textured and on a cluttered/textured table. This gives in total 2*(18*8 + 9*8) = 432 scenes.

All scenes can be viewed here.

The name of the objects can be found here.

Gert Kootstra, Mila Popović, Jimmy Alison Jørgensen, Danica Kragic, Norbert Krüger