search

THU-VCLab / HGGD

Official code of paper "Efficient Heatmap-Guided 6-Dof Grasp Detection in Cluttered Scenes"
MIT License
51 stars 6 forks source link

How to generate grasp labels #3

Closed lishiyu005 closed 6 months ago

lishiyu005 commented 6 months ago

Dear author,

I want to train HGGD for another dataset. I want to know how to generate grasp labels from Graspnet-1billion dataset. Would you like to share your code for grasp generation?

Thank you very much.

Best regards, Shiyu

ChenThree commented 6 months ago

Thanks for your interest in our work! Our data generation pipeline mainly follows the 6-Dof grasp to rect grasp conversion provided in the graspnet dataset api, while the difference is that we save the grasp rotation as Euler angels in the rect grasp class. There may not be an exact schedule for the code release for grasp generation, because it requires extra code reorganization and testing, but if there is any progress in the release of this part, we will update in this repo.

lishiyu005 commented 6 months ago

Thanks for your answer. In my opinion, the current public available dataset for robotic grasping is in different format, which makes the cross training and validation difficult. Therefore it would be great that a unified format can be defined. Thanks for your contribution to simplify grasping labels of Graspnet-1-billion dataset and make the training more efficient anyway.