About the dataset split of the grasp anything++ training and val.

[unira-zwj]

Glad you shared such a great job.

I noticed in the code that the number of split/grasp-anything++/train/seen.obj is 14516, the number of split/grasp-anything++/test/seen.obj is 573, and the number of split/grasp-anything++/test/unseen.obj is 230.

Is these number the correct setting?

[andvg3]

Hi @unira-zwj ,

To some extent, yes. I set the small number for testing. For more comprehensive results, you can increase the number of samples by appending seen.obj and unseen.obj with more idx of the samples.

[Pierre0089]

Hi @unira-zwj, @andvg3

On my side, I found the number of split/grasp-anything++/test/seen.obj is also 573, and the number of seen grasp files is 1376. I made evaluations of GR-ConvNet, and the IOU results is 0.28, not 0.37 in Table 1. What about your results?

[andvg3]

Hello @Pierre0089 ,

The slight drop in results may be due to mismatched dependency versions. We used an older version of PyTorch (1.12); could you please have a check?

[Pierre0089]

Hi,

Thx for your quick feedback.

I have followed the installation https://github.com/Fsoft-AIC/LGD?tab=readme-ov-file#installation. I just checked, and found the installed PyTorch version is 1.12.1.