2D point cloud

[liangzhendong123]

Hello, KPCONv-PyTorch is an excellent project. Does this project support 2d point cloud computing? We hope to promote cloud computing in more fields like digital health and imaging. Thank you very much.

[working12]

2d point cloud means the depth is constant. Theoretically it would be similar to 2d convolution since the input sphere would then become input circle (the points inside which would be fed to the model for training).

[liangzhendong123]

@working12 Thank you very much for your reply. We hope to use KPConv for key point extraction of 2d point cloud, and then match different 2D point aggregation (this can also be understood as transforming 2D image matching into point cloud matching problem). Do you think the depth of different images should be set to different constant values when 2d images are converted into point clouds for key point extraction? Will it affect the feature extraction of KP network?

[HuguesTHOMAS]

Hi @liangzhendong123,

If your 2D point clouds are sparse, using KPConv on your 2D data could make sense.

Theoretically, it is totally possible, that we can define a point kernel in any dimension. However, the current code is not able to handle it. I am working on a new project where I will reimplement everything from scratch and make it usable in any dimension. It will take some time though.