search

Pointcept / PointTransformerV3

[CVPR'24 Oral] Official repository of Point Transformer V3 (PTv3)
MIT License
583 stars 30 forks source link

About estimating surface normals #53

Open boxiao-wv opened 1 month ago

boxiao-wv commented 1 month ago

Thanks a lot for the amazing work! I have a question concerning the surface normals.

I read in the readme that for the dataset S3DIS, the surface normals will help the segmentation results, and it seems that the surface normals are parsed from the meshes of the rooms. Now I have some XYZRGB point clouds but I don't have the mesh. I would like to train PT-V3 on these point clouds, I wonder whether I should first estimate the surface normals and use them as extra inputs for the network, namely, XYZNxNyNzRGB will this likely improve the performance?

Look forward to your reply.

Thank you very much!

Gofinge commented 1 month ago

Hi, although normal is optional in our codebase, based on our experience, normal from mesh (reconstruction) is good for perception. But for raw point clouds like LiDAR (outdoor), which don't contain normal information, we didn't include normal as an additional feature.

boxiao-wv commented 1 month ago

Thanks a lot for the reply! For my the point clouds scans are additional information for example reflectance, return number, and number of returns. Do you think these will help the performance of the network?

Gofinge commented 1 month ago

You can include these characters in a feature and make some adjustments. Usually, additional information won't harm the performance.