search

dvlab-research / Stratified-Transformer

Stratified Transformer for 3D Point Cloud Segmentation (CVPR 2022)
MIT License
376 stars 40 forks source link

How can you apply nn.linear in upsampling #64

Open LEEHK96 opened 2 years ago

LEEHK96 commented 2 years ago

In TransitionUp layer, I just wonder one thing, but I have no idea.

x = self.linear1(x1) + pointops.interpolation(p2, p1, self.linear2(x2), o2, o1)

as you mentioned earlier in different issue, I know what offset do but,

As far as I understand, I think we should divide the number of points per batch in x_2 and apply linear respectively.

like in your work,

> if pxo2 is None:
>             _, x, o = pxo1  # (n, 3), (n, c), (b)
>             x_tmp = []
>             for i in range(o.shape[0]): ....
>

so I wonder how it can be done at once.

I think self.linear2(x2) will combine all points in different data.

If you think I'm wrong, please let me know. Thank you!

and Thank you for great work!

X-Lai commented 2 years ago

I don't really understand your problem. Butx2 contains features for different scenes in the same batch, and all of them share the same linear transformation. I think this is correct.