search

KuangenZhang / ldgcnn

Linked Dynamic Graph CNN: Learning through Point Cloud by Linking Hierarchical Features
MIT License
75 stars 19 forks source link

Difference between the graph feature extraction module and EdgeConv in DGCNN? #4

Open usherbob opened 4 years ago

usherbob commented 4 years ago

Hi Kuangen,

Your paper is really an enlightening work and thanks for sharing your code. But I wonder is there any difference between the graph feature extraction module and EdgeConv in DGCNN. After reading your code, I believe the graph feature extraction function is the same as EdgeConv in DGCNN.

You strengthened that the graph feature extraction module in your work is rotation invariant. And you gave a mathematical explanation through the conversion between (10) and (12) in your paper. This part confuses me a lot, however, would you mind give a detailed explanation about the conversion? image image

KuangenZhang commented 4 years ago

Hi, thanks for your interest. Yes. we used the same EdgeConv layer as in DGCNN. In our paper, we just showed why the EdgeConv itself could extract multi-view features, which could be used to classify the rotated point cloud. However, we cannot ensure that the EdgeConv to be absolutely rotation invariant. We can only say EdgeConv may be approximately rotation invariant if it can extract multi-view features.