Sweep: A Modular Library for Deep Multi-View Stereo Networks.
About
The motivation for this repository is to unify the current Multi-View Stereo network architectures into a single repository using the same network components to allow for a plug-and-play style framework for testing MVS architecture design.
Current Architectures
We are currently supporting the following network architectures:
- MVSNet: Depth Inference for Unstructured Multi-view Stereo | paper
- Non-parametric Depth Distribution Modelling based Depth Inference for Multi-view Stereo | paper
- Generalized Binary Search Network for Highly-Efficient Multi-View Stereo | paper
Current Datasets
We are currently supporting the following datasets (see 'Data Preparation' for downloading each dataset):
Data Preparation
We provide several datasets used in our experiments in an organized and consistent format:
- DTU
- TNT (coming soon...)
- BlendedMVS (coming soon...)
- Replica (coming soon...)
- ScanNet (coming soon...)
If you would like to use your own data, please feel free to create a new class for the dataset in src/datasets/<DATASET>.py. Please feel free to look at any of the existing implementations for reference. They are all very similar. NOTE: The BaseDataset class in the src/dataset/BaseDataset.py file should not need updating to support new datasets. Only the build_dataset function at the top of the file would need updating to be able to locate and import the new dataset. You will receive an error specifying your dataset cannot be found if you forget to update this function. Additionally, please update the choices=["ScanNet", "Replica", "DTU", "TNT"] list for the --dataset argument in the src/config.py file with your new dataset name for argument parsing.
