Wrong results for the grouping_operation. about pointnet2_pytorch HOT 4 CLOSED

I’m happy to take a look, but I need more information than “it’s wrong”. Do you have an input/output example pair?

from pointnet2_pytorch.

Hi @erikwijmans, thank you for your reply. I tested with the following snippets, and found the grouped points 'xyz1' did not match its corresponding points in 'xyz'. I hope you can reproduce it.

xyz_trans = torch.rand((1, 3, 32)).cuda()
idx = torch.cat([torch.range(0,31).view(1, 32, 1),
                 torch.randint(0, 32, (1, 32, 5))], dim=-1).type(torch.long).cuda()
grouped_xyz_src = grouping_operation( xyz_trans, idx)  # (B, 3, npoint, nsample)
print("xyz1: ", grouped_xyz_src[:, :, :, 0])
print("xyz_tran: ", xyz_trans)
print("idxs: ", idx)

from pointnet2_pytorch.

idx should be an int tensor, not a long tensor.

EDIT: Digging into this a little further, I assume you are on the pytorch-1.0.0 branch as the function would have errored out if you had it called it when unexpected types on master. I have added assertions to the pytorch-1.0.0 branch to ensure the tensors are of the expected types.

from pointnet2_pytorch.

Thank you. I got it.

from pointnet2_pytorch.