[paper-review] Dynamic CNN for Learning on Point Clouds
Yue Wang1, Yongbin Sun1, Ziwei Liu2, Sanjay E. Sarma1, Michael M. Bronstein3, Justin M. Solomon1 Salah Rifai1, Pascal Vincent1, Xavier Muller1, Xavier Glorot1, Yoshua Bengio1
Jun. 11
한 문장 요약
Point Cloud의 Feature를 Graph CNN 구조로 추출해보자.
Contribution
- EdgeConv:
- local geometric structure를 얻어냄.
-
edge feature
를 만들어내며, 이는 point와 its neighbor의 relationship을 나타낸다. 그리고 당연하게도 이는 permutation invariant 하다.- 기존의 PointNet++ 에서는 local feature에 대해 포착하지 못하는 한계점을 극복하고자 한다.
- layer를 거쳐가며, dynamic하게 업데이트 되는 grpah에서도 잘 작동한다.
- 쉽게 다른 방법론에 붙일 수 있다.
Approach:
Edge Convolution
The output of EdgeConv at the i-th vertex is given by
\[\begin{equation} \mathbf{x}'_{i}=\square_{k:(i,j)\in \mathcal{E}} h_{\mathcal{\Theta}}(x_i,x_j) \end{equation}\]- \(F\)-dimensional point cloud with \(n\) points; \(\mathbf{X}=\{\mathbf{x}_{1},\cdots,\mathbf{x}_{n}\}\)
- Graph \(\mathcal{G=(V,E)}\) representing local point cloud structure, where \(\mathcal{V}=\{1,\cdots,n\}\) and \(\mathcal{E}\subseteq \mathcal{V\times V}\) are the \(\text{vertices}\) and \(\text{edges}\), respectively.
- Construct \(\mathcal{G}\) as the k-nearnest neighbor (k-NN) graph of \(\mathbf{X}\) in \(\mathbb{R}^{F}\).
- Graph includes self-loop, \(\text{edge~features}\) as \(\mathbf{e}_{ij}=h_{\Theta}(x_i,x_j)\), where \(h_{\Theta}:\mathbb{R}^{F} \times \mathbb{R}^{F} \rightarrow \mathbb{R}^{F'}\)
Dynamic graph update
Properties
Comparison to existing methods
Conclusion:
Thoughts:
Enjoy Reading This Article?
Here are some more articles you might like to read next: