This is part of a study project using the AA-RMVSNet to reconstruct buildings from multiple images.


It is exciting to connect the 2D world with 3D world using Multi-view Stereo(MVS) methods. In this project, we aim to reconstruct several architecture in our campus. Since it’s outdoor reconstruction, We chose to use AA-RMVSNet to do this work for its marvelous performance is outdoor datasets after comparing some similar models such as CasMVSNet and D2HC-RMVSNet. The code is retrieved from here with some modification.


Here we summarize the main steps we took when doing this project. You can reproduce our result after these steps.


First, you need to create a virtual environment and install the necessary dependencies.

conda create -n test python=3.6
conda activate test
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=10.0 -c pytorch
conda install -c conda-forge py-opencv plyfile tensorboardx

Other cuda versions can be found here

Struct from Motion

Camera parameters are required to conduct the MVSNet based methods. Please first download the open source software COLMAP.

The workflow is as follow:

  1. Open the COLMAP, then successively click reconstructionAutomatic reconstruction options.
  2. Select your Workspace folder and Image folder.
  3. (Optional) Unclick Dense model to accelerate the reconstruction procedure.
  4. Click Run.
  5. After the completion of reconstruction, you should be able to see the result of sparse reconstruction as well as position of cameras.(Fig )
  6. Click FileExport model as text. There should be a camera.txt in the output folder, each line represent a photo. In case there are photos that remain mismatched, you should dele these photos and rematch. Repeat this process until all the photos are mathced.
  7. Move the there txts to the sparse folder.



To use AA-RMVSNet to reconstruct the building, please follow the steps listed below.

  1. Clone this repository to a local folder.

  2. The custom testing folder should be placed in the root directory of the cloned folder. This folder should have to subfolders names images and sparse. The images folder is meant to place the photos, and the sparse folder should have the three txt files recording the camera’s parameters.

  3. Find the file list-dtu-test.txt, and write the name of the folder which you wish to be tested.

  4. Run colmap2mvsnet.py by

    python ./sfm/colmap2mvsnet.py --dense_folder name --interval_scale 1.06 --max_d 512

    The parameter dense_folder is compulsory, others being optional. You can also change the default value in the following shells.

  5. When you get the result of the previous step, run the following commands

    sh ./scripts/eval_dtu.sh
    sh ./scripts/fusion_dtu.sh
  6. Then you are should see the output .ply files in the outputs_dtu folder.

Here dtu means the data is organized in the format of DTU dataset.


We reconstructed various spot of out campus. The reconstructed point cloud files is available here (Code: nz1e). You can visualize the file with Meshlab or CloudCompare .


View Github