olyvia_grasping

Grasping

Grasping Dataset for single/cluttered environment

Here we introduced two datasets for grasping- one with single object in a workspace and other with heavily cluttered environment. All the frames have been captured using 3D range device (kinect/ensenso) from a single viewpoint and has point cloud data, annotation, 2D image. Annotation is given in image plane in terms of orientated rectangle. The grasping rectangle format is same as cornell dataset . A handle is represented by 4 lines that give the vertices in a counter-clockwise order and first two coordinates define the line representing the orientation of the gripper plate. The vertices is represented by x and y coordinates separated by a space. For cluttered scene the name of the object is added corresponding to each handle. The source code is uploaded in the gitlab link.

Cluttered Scene-

Consists of 40 frame.
Each fame consists of multiple household objects. Some example images are shown here.
TCS_Dataset_clutter can be downloaded from here.
Annotation Rectangle is given for those objects that can be grasped with two finger in the scene. Annotation of grasping handles is a bit tricky as it depends on physical parameters of the gripper and gap required to pick using the handle. Gripper finger width determines the extent of grasping region parallel to gripper direction. If a complete surface is graspable than we have marked the whole surface with a single rectangle and any rectangle that has same orientation and fits inside that will be considered as true handle. Again, the annotation has been done by assuming small gap around the handle and this gap may vary depending in gripper parameters. Also for some rectangle, it may not possible to perform a grasping action from direction . So, to evaluate the dataset with a different two finger gripper one need to apply some filter on the given annotations.
The grasping rectangle format is same as cornell dataset . A handle is represented by 5 lines. The first line is the name of the object on which the handle is annotated. Next 4 lines represents the four vertices of the ractangle.

Single Object-

This dataset consists of 10 single objects per frame inside a workspace from a single view. A total of 382 frames.
Individual objects includes cleaning_brush, battery, cup, coffee, tissue_paper, dove_soap, fevicol, sprout_brush, clips, toothpaste.
Individual objects can be accessed from here.

Google Sites

Report abuse