Issue fo using my own data

[henry890112]

Hello, I am using my own data to train contact graspnet. However, I am unable to obtain thicker grasping poses in the output. Could this be due to a problem with my data or do I need to modify the test_data/*.npy files? I have attached my input and output data for reference. Thank you!

[MartinSmeyer]

Hey @henry890112,

What robot / gripper model do you use? If your gripper width is wider than the 8cm of the Franka Panda Gripper, then check #28

[MartinSmeyer]

In general thicker grasping poses are a bit less confident, because they are less certain to result in successful grasps. To force such predictions, you can lower the confidence threshold and filter for grasps with higher predicted grasp widths.

[henry890112]

I apologize for the poor description of my issue, but my question is: I am unable to achieve the grasping pose that is highlighted in green, just like which I have attached below. Thank you for your patience.

[MartinSmeyer]

Sorry still unclear, that is not a question

[henry890112]

in my own data I can't get the grasping pose that is highlighted in green, is this Highest Scoring Gripping Pose？and why I can't get it in my own data. Thank you！

[MartinSmeyer]

Yes this is the highest scoring grasp pose. Ah ok, it's just about the visualization, the most confident grasps are only shown if you have more than one object in your mask image:

https://github.com/NVlabs/contact_graspnet/blob/a12b7e3a1d7adf38762bfae1c8b84b4550059a6f/contact_graspnet/visualization_utils.py#L99-L105

[henry890112]

Thank for your answer!