ScanComplete is a data-driven approach which takes an incomplete 3D scan of a scene as input and predicts a complete 3D model, along with per-voxel semantic labels. This work is based on our CVPR'18 paper, ScanComplete: Large-Scale Scene Completion and Semantic Segmentation for 3D Scans.

Training is implemented with TensorFlow. This code is tested under TF1.3 and Python 2.7 on Ubuntu 16.04.

• See run_train.sh for calling the training (will need to provide a path to the train data).
• Trained models: models.zip

• See run_complete_scans_hierarchical.sh for testing on partial scans (needs paths to test data and model).

• Test scenes as TF Records:
  • vox5.zip (4G)
  • vox9.zip (620M)
  • vox19.zip (105M)
• Train data as TF Records:
  • vox19_dim32.zip (12G) for training the 19cm hierarchy level
  • vox5-9-19_dim32.zip (240G) for training the 9cm and 5cm hierarchy levels. IMPORTANT: For training a hierarchy with more than just the finest level (e.g., 19-9-5 or 9-5 instead of just 5cm), the finer levels should be trained using the results from the trained model from the previous hierarchy level; i.e. this data will need to be edited.

If you find our work useful in your research, please consider citing:

@inproceedings{dai2018scancomplete,
  title={ScanComplete: Large-Scale Scene Completion and Semantic Segmentation for 3D Scans},
  author={Dai, Angela and Ritchie, Daniel and Bokeloh, Martin and Reed, Scott and Sturm, J{\"u}rgen and Nie{\ss}ner, Matthias},
  booktitle = {Proc. Computer Vision and Pattern Recognition (CVPR), IEEE},
  year = {2018}
}

If you have any questions, please email Angela Dai at [email protected].