how about directly using STN to predict transformation?

[zhengzh]

Hi, I read your paper, but i can't understand the benefit of using LK algorithm, how about directly using PointNet with source and target point clouds as input to predict transformation and use the transformation to transform source point cloud, then minimize the Chamfer distance loss between transformed source point cloud and target point cloud.

Thanks

[rarunsrivatsan]

We have tried doing a version of what you mentioned. What happens is the resulting network produces results comparable to PointNetLK for objects it was trained on, but for other categories it produces bad results. The advantage of PointNetLK is that we use learning to only understand the feature space, but use mathematically established techniques to perform the alignment. As a result the alignment part is not dependent on the shapes that the network was trained on, making it more robust. However, if you application requires only a fixed set of objects whose pose you need to find, you can totally directly estimate the pose using chamfer distance or other such variants. Thanks for your interest in our work.

[zhengzh]

Oh, I get it. Thanks for your response!