search

Yochengliu / Relation-Shape-CNN

Relation-Shape Convolutional Neural Network for Point Cloud Analysis (CVPR 2019 Oral & Best paper finalist)
https://yochengliu.github.io/Relation-Shape-CNN/
MIT License
415 stars 71 forks source link

Confused about the symmetric aggregation function #3

Closed zengjinghui closed 5 years ago

zengjinghui commented 5 years ago

dear author, RS-CNN is excellent! but I am confused about the aggregation function, which you said it's symmetric, such as summation or max pooling, in Session 4.3, Figure 8, in the 1st layer, there are 512 points(1/2 of the input points), but in the 2nd layer, there remain only 128 points(1/4 of the 1st layer), why? the relation-shape convolution(implemented as shared MLP) and channel-raising mapping(also shared MLP) don't seem to change the point number, because of the aggregation function? how? Thank you! Looking forward to your reply!

Yochengliu commented 5 years ago

Sorry for confusion. The change of points' number is achieved by sampling. You can sample 512 points from the input 1024 points, and then take the 512 points as the local centroids x_i to construct local point subsets N(x_i) for "convolution". Please refer to the 2nd part in Sec 3.5 and Fig. 4(b). If it is still difficult for you to understand this part, you can refer to the codes of PointNet++, or wait for our codes.

zengjinghui commented 5 years ago

Oh! Thank you! I think I understand now!