search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
523 stars 114 forks source link

dgcnn #41

Open Manojbhat09 opened 3 years ago

Manojbhat09 commented 3 years ago

Hello Weijing, again your work is amazing for point based approach over grid based approaches. Thank you.

There is a similar approach for graph based methods aka. EdgeConv. I tried to point out some differences between this method and DGCNN, but it looked mostly similar but with Relu activations after extracting edge features. If you have given it a read, would it be possible to point out some overview method that is different?

If the method is vastly different, would these other ways of extracting features from edges and points provide a much better performance or even a difference?

Thanks again for the teachings!

zye1996 commented 3 years ago

Hello Weijing, again your work is amazing for point based approach over grid based approaches. Thank you.

There is a similar approach for graph based methods aka. EdgeConv. I tried to point out some differences between this method and DGCNN, but it looked mostly similar but with Relu activations after extracting edge features. If you have given it a read, would it be possible to point out some overview method that is different?

If the method is vastly different, would these other ways of extracting features from edges and points provide a much better performance or even a difference?

Thanks again for the teachings!

Hi I am also looking into those papers so I probably could give some notes here if it is not too late:

  1. DGCNN uses features to construct a graph and it is dynamic at each layer. The graph is constructed using KNN so it is unidirectional while Point-GNN uses ball query so the graph is bidirectional. For dgcnn, it does not carry the global location of each point to the next layer while for Point-GNN it has to as the graph relies on global location.
  2. The conv operations are quite similar though Point-GNN is modelling against the residual at each update and DGCNN is modelling against the whole feature vector at each node