PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

[@guruprasadsomasundaram9273]

I remember this being such a landmark achievement in 3D computer vision. I used to work with point clouds and meshes and this technique just opened a lot of doors for us.

[@qudezheng]

Brilliant paper! While 2D image classification is more or less a solved problem. There're lots of exciting things happening in the 3D space

[@tksl9236]

读了这个paper，感觉很有创新性

[@jimmyshenmusic]

This work is cool and it creates a new world for the 3D vision understanding. Surprisingly, it even did not use the CNN.

[@gledsonmelotti7497]

Hello, could you tell me why this PointCloud method is better than using a 3D CNN (advantages and disadvantages)?

[@jimmyshenmusic]

3D CNN is computation expensive. We do have some works are using 3D CNN. You can check the leader board of the ModelNet 40. Currently, we have method: based on voxel(typical using 3D CNN), mesh(seems not so popular), raw point cloud(it is becoming popular) and project 3D image to 2D then using multi-view 2D CNN. We can not say which one is better. From the lead board, it seems that the multi-view 2D CNN can get the best result. However, the approach based on point cloud is interesting as in the RGB-D and outdoor LiDAR sensor, the data representation is either point cloud itself or very friendly to raw point cloud. That is one of the most important reasons that this work is interesting. Another important reason is that this work shows some theoretical analysis which is very useful to help us have a deeper understanding about their framework.

[@gledsonmelotti7497]

Hi, how are you? Thank you very much for the explanations about my doubts. Thank you very much. His explanations are very interesting and useful.

[@mattiafucili7597]

What do you mean by "it even did not use the CNN."? If you look at the implementation code in TensorFlow they use a 2Dconv layer everywhere before the max pooling operation. So, the problem is mine and I can't understand the code or did you say that because in the image of the net they wrote MLP instead of 2Dconv?

[@jimmyshenmusic]

Mattia Fucili To be honest, I didn't look the code. Based on the paper, the author mentioned only FC is used.

[@erfanmoshiri3117]

such a string and backed theory! probably will be a lead to many other important projects

[@quantummath]

Great Paper! I'm using it for a project I'm currently taking part in.

[@noemi9351]

Hi, I am starting to use Point Net for a project of mine as well and just saw your comment here. Do you have any information available online about your project, e.g. Research Gate? I'm curious to see all the different projects with PN at the moment. Best

[@chillfuturist9668]

@@noemi9351 I am working on a project using this as well. Are you still working in this space?

[@endingalaporte]

a new wide range of applications open, ahead of us

[@JeracVideo]

Hi, I'm searching for software that can detect eyes, nose, and ears on a human face in a point cloud and return the coordinates of those features. Then, I'd like to merge a second point cloud according to these coordinates. I'm not sure where to start. Can someone give me some hints? Thanks!

[@yifeipeng389]

Does the model takes labels to do segmentation?

[@金属光泽]

感觉这两篇paper的引用爆炸了

[@jobssteve2920]

还有一篇是什么？

[@shindepratibha31]

Brilliant paper and work. I am thinking of incorporating rgb in pointnet++ model. How can I do that? Could you please help with some suggestions?

[@JamesDear-b3q]

did you manage to figure this out? it would greatly help me thanks

[@sanjanakambar2947]

What happens if we add Global Average Pooling after feature transform?

[@vivekgr3001]

Excellent Explanation!

[@gracebear]

Go, Charles Qi ! :)

[@charlieli7486]

Great work!

[@kalpanashowroom2576]

What is the meaning of critical points?

[@bindukarki5616]

Not entirely sure but I think those are the points that are distinct and best represents/distinguish the objects OR in another word key-points in point cloud library... check this: stanford.edu/~rqi/pointnet/ or check for key points in point cloud library for some understanding..apart from being distinct key-points and critical points might be same

[@Rorypandanda]

Yeah, I believe those are the points most responsible for influencing the classification decision of the network.

[@jimmyshenmusic]

The critical points are inferred from the learned global features after the max pooling. From the paper, the critical points will provide the main contribution on the classification task. From the vis result, critical points can well represent the geometric shape of the cloud points.

[@gledsonmelotti7497]

Hello, but how to get the critical points?

[@jimmyshenmusic]

traverse all the raw cloud points. Check whether it will give a contribution to the max pooling result, if it does, then it is a critical point, if not, then it is not a critical point.

[@operaworker]

It's truly a great paper and great explanation, however the chomping sound, which is made by the speaker at the end of each sentence, is horribly disgusting and annoying.

[@Imandabbagh]

I didn't notice until reading your comment, now I can't focus on the topic anymore!

[@saisachinnekkanti7763]

@@Imandabbagh ;(

[@williambonvini5806]

"horribly disgusting"? Don't be so dramatic please 😅 great paper anyways!