Question about Shape Completion

[pwang649]

Fantastic work!

I wonder how to adapt this work to the task of point cloud completion. E.g., what should the position embedding be after the encoded student feature step? How to convert the codebook vectors back into point cloud? Would be great if you can give more details on that!

Are you also thinking of open sourcing the code for point cloud completion?

Many thanks!

[LAB123-tech]

Hello, thank you very much for your recognition of our work. Since works like PointBERT and PointMAE do not provide downstream tasks for point cloud completion,

when working on this downstream task, we referred to the pipeline of the paper "PoinTr: Diverse Point Cloud Completion with Geometry-Aware Transformers". Specifically, we directly input the incomplete point cloud into the network and finally use FoldingNet to restore the original shape of the point cloud. Remember to replaced the encoder with the pre-trained encoder, while keeping the position embedding settings the same as in PoinTr.

As for the vectors in the codebook, in this paper, they represent the cluster centers of geometric point cloud patch vectors. We do not need to convert the codebook vectors back into point clouds.