Clarification on Model Input: How to Handle x and xyz When Using Only xyz

[Monkey-D-Luffy-only]

Hi,

Thank you for sharing this excellent repository! I am currently using your code for classification tasks, where the Model requires both x and xyz as inputs, as shown in the example:

model = Model(in_dim=6, out_dim=40).cuda()
xyz = torch.randn(2, 3, 1024).cuda()
rgb = torch.rand(2, 3, 1024).cuda()
x = torch.cat([xyz, rgb], dim=1)
out = model(x, xyz)
print(out.shape)  # (2, 40)

However, in my case, I only have the xyz coordinates of the point cloud (no RGB or other additional features). I would like to confirm:

• How should x and xyz be set in this scenario?

• If x is set to xyz directly, do I need to adjust in_dim when initializing the model?

• Are there any potential limitations or things to be aware of when only using xyz as input?

Thank you for your time and help! Looking forward to your clarification.

[kentechx]

Assigning xyz to x is just fine.

model = Model(in_dim=3, out_dim=40).cuda()
xyz = torch.randn(2, 3, 1024).cuda()
x = xyz
out = model(x, xyz)
print(out.shape)  # (2, 40)

However, xyz is not rotation-invariant, making it hard to fit the training data for all possible rotations. To alleviate this issue, better preprocess the input point cloud, e.g. aligning the point cloud to a canonical space.

[Monkey-D-Luffy-only]

Assigning xyz to x is just fine.

model = Model(in_dim=3, out_dim=40).cuda()
xyz = torch.randn(2, 3, 1024).cuda()
x = xyz
out = model(x, xyz)
print(out.shape)  # (2, 40)

However, xyz is not rotation-invariant, making it hard to fit the training data for all possible rotations. To alleviate this issue, better preprocess the input point cloud, e.g. aligning the point cloud to a canonical space.

Thank you for your response! Since I haven’t carefully reviewed the code, I’d like to ask if the model framework and default settings are fully consistent with the original paper and the official repository (https://github.com/guochengqian/PointNeXt)?

[kentechx]

Not fully consistent, but the differences in implementation are minimal. FYI, the performance in my experiment is as follows.(https://github.com/kentechx/s3dis_sem_experiments?tab=readme-ov-file#results)

Model	input	6-fold mIoU (%)	6-fold OA (%)	area 5 mIoU (%)	area 5 OA (%)	device	report
PointNext-S	rgb+height	67.5	88.1	65.2	89.0	1x 3090	report
PointNext-XL	rgb+height	71.4	89.2	68.2	90.1	1x 3090	report

[Monkey-D-Luffy-only]

Not fully consistent, but the differences in implementation are minimal. FYI, the performance in my experiment is as follows.(https://github.com/kentechx/s3dis_sem_experiments?tab=readme-ov-file#results)

Model input 6-fold
mIoU (%) 6-fold
OA (%) area 5
mIoU (%) area 5
OA (%) device report
PointNext-S rgb+height 67.5 88.1 65.2 89.0 1x 3090 report
PointNext-XL rgb+height 71.4 89.2 68.2 90.1 1x 3090 report

OK, thanks.