About the low quality of grasping pose

[lsattf]

The algorithm cannot detect the objects in the scene very well. But the paper mentioned it has the object detection or segmantation algorithm included. Do I need to do the object detection in point cloud myself before feed into the package or I used the package incorrectly?

Here is the test result from a .pcd file with some objects on the table, you can see it just detect one or two of the obejcts and generate some pose on wall and table legs:

[atenpas]

Yes, you need to do the object detection method. GPD's purpose is to detect grasp poses, not objects.