search

charlesq34 / pointnet2

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
Other
3.13k stars 900 forks source link

Where can I find T-net module #35

Open itswyri opened 6 years ago

itswyri commented 6 years ago

Hello,

It seems that there is no T-net module in function "pointnet_sa_module" in utils/pointnet_util.py, Is there no T-net module in this code? or am I missing something? Please let me know where I can find it.

CrazySnailer commented 6 years ago

@itswyri I think the author didn't use the T-net in the pointnet_sa_module.

AIMan-Zzx commented 5 years ago

t-net and max pooling is the key of pointnet, if pointnet++ did not use above ,why papers did not mark

adioshun commented 5 years ago

@charlesq34 could you give me the commet regarding the @itswyri 's question? I also wondering the T-Net usage in PointNet++.

AIMan-Zzx commented 5 years ago

no t-net is right ，you can read the original paper

发自我的iPhone

------------------ Original ------------------ From: adioshun notifications@github.com Date: Tue,Aug 27,2019 8:47 AM To: charlesq34/pointnet2 pointnet2@noreply.github.com Cc: sunnyForIOS 276749553@qq.com, Comment comment@noreply.github.com Subject: Re: [charlesq34/pointnet2] Where can I find T-net module (#35)

@charlesq34 could you give me the commet regarding the @itswyri 's question? I also wondering the T-Net usage in PointNet++.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

croros commented 5 years ago

For anyone still wondering about this, see this line from the paper: "Note that PoinetNet (vanilla)... is the version... that does not use transformation networks, which is equivalent to our hierarchical net with only one level" Notice the wording in PointNet++: the "PointNet layer uses a mini-PointNet" i.e. not a normal PointNet. As a matter of fact, in the original PointNet paper T-Nets are referred to as mini-PointNets (they're not exactly the same, but similar).

As for why they decide use a vanilla PointNet instead of a normal one, maybe it was for efficiency reasons as they mention it is less than half the FLOPS/sample as the normal PointNet.