search

lijx10 / USIP

USIP: Unsupervised Stable Interest Point Detection from 3D Point Clouds. ICCV 2019
GNU General Public License v3.0
202 stars 28 forks source link

Convert the point cloud "ply" to "npy" #24

Open tuanho27 opened 3 years ago

tuanho27 commented 3 years ago

Hi there,

Thanks for the great work. I've tried to inference my own point cloud based on your 3D match pretrained model and network. But I met a problem that the weight asking for the input to be shaped at Nx7 while the point shape is just Nx3.

Even I took a look at the evaluation data loader (match3d_eval_loader.py), still confusing. I just wonder that how did you convert the point cloud from ply to npy? Because when I check the 3D_Match_eval_npy, it's already converted to shape Nx7. And Is the surface_normal_len need for inference?

Thank you in advance, Tuan

lijx10 commented 3 years ago

Hi,

The other 4 elements are surface normal (nx, ny, nz) and curvature (scalar). They are computed using PCL library. In fact, these 4 elements has almost no effect to the performance. You may simply train your own model using Nx3. Thanks.

Best Regards Jiaxin Li

tuanho27 notifications@github.com 于2021年1月10日周日 下午4:37写道：

Hi there,

Thanks for the great work. I've tried to inference my own point cloud based on your 3D match pretrained model and network. But I met a problem that the weight asking for the input to be shaped at Nx7 while the point shape is just Nx3.

Even I took a look at the evaluation data loader (match3d_eval_loader.py), still confusing. I just wonder that how did you convert the point cloud from ply to npy? Because when I check the 3D_Match_eval_npy, it's already converted to shape Nx7. And Is the surface_normal_len need for inference?

Thank you in advance, Tuan

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lijx10/USIP/issues/24, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC5G72PVP5ACA24H5MNVKCDSZFRNBANCNFSM4V4IALRA .

-- 黎嘉信 Jiaxin LI PhD, National University of Singapore B.S., Tsinghua University Mobile: 65-9094 1909 / 86-15201519053 Email: lijx1992@gmail.com / jli@u.nus.edu jli@nus.edu.sg

tuanho27 commented 3 years ago

Hi, thanks for the reply.

So, It seems that I cannot directly use the provided pretrained model to test my point cloud. PCL is the great lib, but I may dig into it later. I think I will try to re-train the model without the 4 added elements if it does not very effective to the performance.

Regards, Tuan