search

rabbityl / DeformationPyramid

[NeurIPS 2022] Non-rigid Point Cloud Registration with Neural Deformation Pyramid
MIT License
178 stars 23 forks source link

Are the two data sets required the same? #10

Open balabalaboy opened 1 year ago

balabalaboy commented 1 year ago

Hello, I am a beginner at point cloud non-rigid registration, thank you for your outstanding work. I saw your Lepard work on the web, followed by LNDP on your GitHub, and it looks like it's an updated version of Lepard that works better than Lepard. After checking your operation requirements, I would like to ask whether the 4Dmatch data link in Lepard and the 4Dmatch data link in LNDP are the same. They look the same size and name, but the google drive size is somewhat different. Because my Google Drive download speed is slow (only 2b/s) can I download only one of them if I want to re-run both jobs or do I need to download both linked datasets? In addition, is it convenient for you to provide the OneDrive version? https://github.com/rabbityl/lepard#download-data-and-pretrained-model

In addition, I would like to ask you whether you can achieve global and local registration. Like a whole body point cloud image and a belly scan image or a face scan image?

rabbityl commented 1 year ago

Hi, in the NDP paper, we use filtered 4DMatch benchamrk, which removes point cloud pairs with near-rigid movement. I.e. the filtered benchmark is smaller than the original one in Lepard paper. We recommed using the filetered 4DMatch for valid non-rigid registration evaluation. You could find the details in the last page of the NDP paper.

For whole-to-part registration, yes. you can do it using NDP with a little proper parameter tuning effort.

balabalaboy commented 1 year ago

Hi, in the NDP paper, we use filtered 4DMatch benchamrk, which removes point cloud pairs with near-rigid movement. I.e. the filtered benchmark is smaller than the original one in Lepard paper. We recommed using the filetered 4DMatch for valid non-rigid registration evaluation. You could find the details in the last page of the NDP paper.

For whole-to-part registration, yes. you can do it using NDP with a little proper parameter tuning effort.

Thank you for your prompt reply.❥(^_-) I will try. In addition, I would like to ask whether the NDP can still Handel the point cloud even if the source of the point cloud is different, such as the point cloud from Lidar and the point cloud from 2D image reconstruction, similar to smpl.