no, n will not be fixed. he mentiones specifically that each point is fed through the MLPs individually. That means it doesn't matter how many points (n) you have to begin with. Also at the end, the max pool operates over all point = again it doesn't matter how many there are to begin with. 1024 is the amount of "features" that are extracted for each pointcloud
It allows the network to find more standard representations of points clouds, which reduces the network complexity required downstream of the transformation layer to generalize to certain transformations (for example shape rotations). Basically: you input rotated airplane, network unrotates airplane, network doesn't need to learn hypercomplex conditional rules in further layers.
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
I am a student from an international university and I must say that this professor is amazing! He explains really well!
Super clear explanation! Thanks alot
Great work. Thank you
Great explanation. Thank you alot.
Very clear thanks !
please ... I didn't understand how MLP, input transform and feature transform work ? please answer me
n is fixed ?
yes 1024
no, n will not be fixed. he mentiones specifically that each point is fed through the MLPs individually. That means it doesn't matter how many points (n) you have to begin with. Also at the end, the max pool operates over all point = again it doesn't matter how many there are to begin with. 1024 is the amount of "features" that are extracted for each pointcloud
What is the need of transformation, anyway the point cloud won't loose it's structure before and after
It allows the network to find more standard representations of points clouds, which reduces the network complexity required downstream of the transformation layer to generalize to certain transformations (for example shape rotations).
Basically: you input rotated airplane, network unrotates airplane, network doesn't need to learn hypercomplex conditional rules in further layers.