3D Building Reconstruction

[gogogo617]

Hello, thank you for sharing such wonderful work!
However, there is a problem that is bothering me. It seems that I have not found the steps to rebuild and visualize the building. Can you help me solve it? Thank you！

[zhenghaohu]

Thank you for your attention. In this part of the post-processing, we assign the predicted height values to the pixels of the predicted building masks, then convert the raster result to vector shapefile. Finally, we input the satellite images and prediction results together into ArcScene 10.8 for visualization.

[zhangying2576]

感谢您的关注。在这部分后处理中，我们将预测的高度值分配给预测的建筑物掩膜的像素，然后将栅格结果转换为矢量 shapefile。最后，我们将卫星图像和预测结果一起输入到 ArcScene 10.8 中进行可视化。

Should the network be retrained to achieve building segmentation, or can the prediction results be processed directly? Could you please introduce the specific process of realizing three - dimensional visualization for the prediction results of this network? Thank you.

[zhenghaohu]

感谢您的关注。在这部分后处理中，我们将预测的高度值分配给预测的建筑物掩膜的像素，然后将栅格结果转换为矢量 shapefile。最后，我们将卫星图像和预测结果一起输入到 ArcScene 10.8 中进行可视化。

Should the network be retrained to achieve building segmentation, or can the prediction results be processed directly? Could you please introduce the specific process of realizing three - dimensional visualization for the prediction results of this network? Thank you.

The output of the model includes the segmentation mask results of the building, which do not require retraining. We vectorize the segmentation mask results and use the height prediction results as attributes of the vector polygon. With software, we can obtain 3D visualization results based on the height attribute values.

[zhangying2576]

感谢您的关注。在这部分后处理中，我们将预测的高度值分配给预测的建筑物掩膜的像素，然后将栅格结果转换为矢量 shapefile。最后，我们将卫星图像和预测结果一起输入到 ArcScene 10.8 中进行可视化。

应该重新训练网络以实现建筑物分割，还是可以直接处理预测结果？能否介绍一下这个网络的预测结果实现三维可视化的具体过程？谢谢。

模型的输出包括建筑物的分割掩码结果，这些结果不需要重新训练。我们对分割掩码结果进行矢量化处理，并将高度预测结果用作矢量多边形的属性。借助软件，我们可以根据高度属性值获得 3D 可视化结果。

I noticed in the paper that the segmentation results are regularized and polygonized, with a reference to [38]. Has the implementation code been open-sourced? I am just starting to work on 3D reconstruction and would like to ask what method is used for vectorizing segmentation masks? Looking forward to your reply, thank you very much.

[zhenghaohu]

感谢您的关注。在这部分后处理中，我们将预测的高度值分配给预测的建筑物掩膜的像素，然后将栅格结果转换为矢量 shapefile。最后，我们将卫星图像和预测结果一起输入到 ArcScene 10.8 中进行可视化。

应该重新训练网络以实现建筑物分割，还是可以直接处理预测结果？能否介绍一下这个网络的预测结果实现三维可视化的具体过程？谢谢。

模型的输出包括建筑物的分割掩码结果，这些结果不需要重新训练。我们对分割掩码结果进行矢量化处理，并将高度预测结果用作矢量多边形的属性。借助软件，我们可以根据高度属性值获得 3D 可视化结果。

I noticed in the paper that the segmentation results are regularized and polygonized, with a reference to [38]. Has the implementation code been open-sourced? I am just starting to work on 3D reconstruction and would like to ask what method is used for vectorizing segmentation masks? Looking forward to your reply, thank you very much.

We used the model from [38] for regularization (this step is not necessary), please refer to link for details. Of course, you can also use other better methods or models. Vectorization is implemented using libraries such as rasterio and geopandas.