search

nv-tlabs / NKSR

[CVPR 2023 Highlight] Neural Kernel Surface Reconstruction
https://research.nvidia.com/labs/toronto-ai/NKSR
Other
765 stars 44 forks source link

The results of my test seem to be less than ideal. #10

Closed xiemeilong closed 1 year ago

xiemeilong commented 1 year ago

The results of my test seem to be less than ideal. The normals in my point cloud were generated in CloudCompare, and I'm not sure if that's the reason. image point cloud image generated normals image

heiwang1997 commented 1 year ago

Thanks for sharing the result! @xiemeilong

The normal you've estimated seems fine. We don't explicitly tackle texture in our paper, if you want higher-quality texture, you can try increasing mise_iter (e.g. to 2) during mesh extraction.

Also, I guess our training data doesn't contain construction sites, and you can try to fine-tune the model if you have data containing tubes/scaffolds.

Happy to discuss more if you are concerned with any specific issues in your reconstruction.

fwilliams commented 1 year ago

@xiemeilong Could you upload a copy of your point cloud with normals and colors (ideally in PLY)?

xiemeilong commented 1 year ago

@heiwang1997 Due to an error in executing "vis.show_3d" on my computer, I saved the model using "vis.to_file" and rendered it in Blender instead. Since there were no textures, the rendering was done using vertex colors.

xiemeilong commented 1 year ago

@fwilliams I'm sorry, the point cloud belongs to the company and it is a power facility, so it is not suitable for be public .

heiwang1997 commented 1 year ago

I see. We really hope to further improve NKSR in the future with the feedback from the community, but just saying 'less than ideal' is a bit vague for us to make active changes. Would you please kindly point us to some possible suggestions for improvement?

heiwang1997 commented 1 year ago

Closing due to inactivity. Feel free to re-open if you still have a problem.