关于方法实时性和稀疏点云问题咨询

[yakexi-cmd]

作者您好，您的方法给了我很大的启发，我想咨询一下该方法是否可以实时构建分割和重建结果？以及如果我的输入点云比较稀疏和不完整，是否可以补全缺失的部分点云？

[chenzhaiyu]

Thanks for your interest!

• Efficiency & real-time: Occupancy inference and surface extraction are fast, but the space-partitioning step (which generates polyhedral candidates) can be time-consuming, especially for complex buildings with many planar primitives. The current pipeline is not real-time. For large-scale training/inference, you can check PolyGNN, which is generally more efficient.

• Incomplete inputs: Our paper assumes that the major planar primitives can be extracted. For recovering primitives from sparse or incomplete point clouds, please check our follow-up work PaCo, which addresses parametric completion and provides compatiable inputs for Points2Poly/PolyGNN.

[yakexi-cmd]

非常感谢您的回复！还想请教一下以下问题：我的原始点云是pcd文件，我已经将它转换到了obj文件，之后需要怎么将其转换到vg格式呢？是否有文件代码可以批量处理obj到vg文件的转换？我下载了easy3d，但其中并没有mapple。如果得不到vg文件，是不是无法将obj用于您的reconstruction.py进行后续处理？

[chenzhaiyu]

Indeed you'll need VG files to proceed. You can download Mapple here which provides the GUI. For batch processing, you can use Easy3D with C++/Python, or the GoCoPP CLI.

[yakexi-cmd]

好滴好滴，谢谢您的回复，我去尝试一下相关工作，非常感谢！！

[yakexi-cmd]

作者您好，还需要再打扰您一下，我使用您的point2poly代码运行python reconstruct.py dataset_name='helsinki_my' model_name='helsinki_fullview'（其中helsinki_my是我自己的数据集文件，其中存放了vg合适的表面提取内容），最终在我的输出中query_pts_ms_vis和vis存储了ply格式的点云数据，我使用meshlab打开后得到的是点云文件，似乎这两个文件与原始点云的不同在于进行了点云补全？（希望您能解答一下这样理解是否正确？），我似乎并没有得到多面体的图形，您能帮我解答一下问题出现在哪里吗？

[chenzhaiyu]

This project does not involve point cloud completion. In case no output is generated this is likely because your data is very different from the training data. Please also refer to those previous issues.