Inputs of KPConv

[BingHan0458]

Hi Hugues, Firstly thanks again for releasing your code and some pretrained models. I've been working on this for a few days trying to see if it was possible to use your convolution operator "KPConv" in my own research (for the point cloud upsampling task). But there is a question when I want to reproduce this code -- "class KPConv(nn.Module)" in https://github.com/HuguesTHOMAS/KPConv-PyTorch/blob/master/models/blocks.py, that is: what is the meaning of these inputs -- "q_pts, s_pts, neighb_inds, x"? I can't find the relevant description. Moreover, what I want to know is whether KPConv is plug and play? If I only have point cloud features and neighbors, can I use it for feature extraction? Looking forward to your reply. Best Regards.

[HuguesTHOMAS]

Hi @BingHan0458,

what is the meaning of these inputs -- "q_pts, s_pts, neighb_inds, x"? I can't find the relevant description.

        q_pts (Tensor): query points (M, 3).
        s_pts (Tensor): support points carrying input features (N, 3).
        neighb_inds (LongTensor): neighbor indices of query points among support points (M, H).
        x (Tensor): input features values (N, C_in).

Moreover, what I want to know is whether KPConv is plug and play? If I only have point cloud features and neighbors, can I use it for feature extraction?

Yes, it is. However, in the current implementation, it will be quite hard to use the operation outside the whole pipeline. I am currently working on a new implementation that is more plug and play and where the KPConv operator is more standalone, easier to be used in other types of network.