search

gilbaz / LORAX

CVPR Paper - "3D Point Cloud Registration for Localization using a Deep Neural Network Auto-Encoder" - Partial Implementation Code
91 stars 29 forks source link

register two skull point clouds #3

Open LUOmengce opened 6 years ago

LUOmengce commented 6 years ago

Using RSCS + FPFH to register the two skull point clouds, how to set the parameters of the RSCS and why the registration result is particularly poor. There are 100,000 points in skull point cloud, and it is related to the random point set division

gilbaz commented 6 years ago

Hello,

So you really need to set the division so that it captures semantically recognizable sections. It is quite hard to register just generic sections. Each dataset is different in this aspect. I would try to calculate the FPFH histogram by the center of each SuperPoint to make sure that the radius of the FPFH is smaller than the SP radius. With the RSCS I would create a few full coverage iterations the SP are randomly chosen each time making it pretty robust to changes. Again it is important that you make sure that the RSCS has a radius that captures registration level features.

Let me know how it goes. Different for each dataset. The FPFH really wasn’t that amazing ..

-Gil

On 26 Nov 2017, at 13:25, LUOmengce notifications@github.com wrote:

Using RSCS + FPFH to register the two skull point clouds, how to set the parameters of the RSCS and why the registration result is particularly poor. There are 100,000 points in skull point cloud, and it is related to the random point set division

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

LUOmengce commented 6 years ago

Hello, These days, I still do not set the parameters, resulting in ISS + FPFH registration results are far better than RSCS + FPFH method, but also with less time. Is RSCS not suitable for small-scale point cloud. When you travel back, if you have time, can you set the parameters for me? Originally wanted to send my skull cloud data to you, but did not know your email address. Thanks again for your help!

gilbaz commented 6 years ago

Hi,

Yes this is designed for large-scale point clouds. If you want send me your point cloud and I will take a look.

Thanks, Gil

On 29 Nov 2017, at 13:36, code notifications@github.com wrote:

Hello, These days, I still do not set the parameters, resulting in ISS + FPFH registration results are far better than RSCS + FPFH method, but also with less time. Is RSCS not suitable for small-scale point cloud. When you travel back, if you have time, can you set the parameters for me? Originally wanted to send my skull cloud data to you, but did not know your email address. Thanks again for your help!

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

LUOmengce commented 6 years ago

Thank you for your prompt reply！ I don't use GitHub to send documents very much. Would you please email me? I send email to you, what format of the point cloud file you are most convenient？

gilbaz commented 6 years ago

As written in the paper: gil.elbaz@alumni.technion.ac.il

LUOmengce commented 6 years ago

Hello! Sorry to bother you again！Your papers section 2.3 and 2.7 of the code, due to my limited ability to really achieve it, if you want, can you send me a code to let me learn? Or only need to convert the point cloud code into a depth map to me? thank you for your help

siliangwu commented 4 years ago

Hello, These days, I still do not set the parameters, resulting in ISS + FPFH registration results are far better than RSCS + FPFH method, but also with less time. Is RSCS not suitable for small-scale point cloud. When you travel back, if you have time, can you set the parameters for me? Originally wanted to send my skull cloud data to you, but did not know your email address. Thanks again for your help!

Hello, whether it is convenient to provide the source code of RSCS + FPFH or ISS + FPFH for me to learn, thank you very much!