This repository is a fork of the Stanford 3D Object Point Cloud Completion Benchmark that implements various deep learning methods for point cloud completion including PCN, TopNet, and FoldingNet.
Note: This repository only implements the baseline methods which are PCN, TopNet, and FoldingNet. Please visit our main repository Point Cloud Scene Completion of Obstructed Building Facades with Generative Adversarial Inpainting for code to perform scene completion using generative adversarial inpainting.
Get the original building facade dataset with 11 scenes (input + ground truth) here.
To use the building facade dataset processed according to the Completion3D format, download and extract the zip file into the data directory. The dataset consists of 467 training and 11 validation point cloud pairs stored in H5 format, with each point cloud containing 16384 points. The point cloud pairs are made up of a partial point cloud (stored under the partial directory) and a complete point cloud (stored under the gt directory).
Follow the instructions to setup a Tensorflow Python environment. Next, run the script train_facade.sh to start training. The script will train multiple networks with the building facade dataset according to a cross-validation scheme (i.e. 10 scenes are used for training and 1 scene is used for validation for each of the 11 scenes). The trained models and predictions will be saved in the tensorflow/results/facade directory. Modify the NET parameter in the training script to train with different network architectures including PCN, TopNet, and FoldingNet.
Run the Python script convertH5ToPLY.py to convert the H5 result files to the PLY format that will be used for evaluation. The predicted point clouds will be rescaled and transformed back to the coordinate system of the original input point clouds. The fix_color.py script is used internally to assign color channels to the predicted point cloud based on nearest neighbor search from the input point cloud.
Use the get_accuracy.py script to compute evaluation metrics such as voxel precision, voxel recall, F1-score, position RMSE and color RMSE. Note that the ground truth point cloud should come from the original building facade dataset (before conversion to the Completion3D format).
python get_accuracy.py data/facade_original/01_mason_east_gt.ply baselines/PCN/01_mason_east.ply
The following directories contain the result point clouds after training with the building facade dataset (Upsampled refers to the case where the number of points is increased to 1000000 using mesh-based resampling).