NeuralMVS: Bridging Multi-View Stereo and Novel View Synthesis
Radu Alexandru Rosu 1,
Sven Behnke 1
1University of Bonn, Autonomous Intelligent Systems
This is the official PyTorch implementation of NeuralMVS: Bridging Multi-View Stereo and Novel View Synthesis
NeuralMVS can process RGB images of a scene in order to perform novel view rendering and additionally learns the 3D structure of the scene in an unsupervized way. It can render novel high resolution images in real-time together with the depth of the scene and a confidence map. The implementation is written in PyTorch.
For NeuralMVS we use DataLoaders for loading the data and interface it using EasyPBR. The dependencies of these two packages and also NeuralMVS can be installed with:
$ sudo apt-get install python3-dev python3-pip python3-setuptools
$ sudo apt-get install libglfw3-dev libboost-dev libeigen3-dev libopencv-dev
$ sudo python3 -m pip install piq tensorboardThe easiest way is to sequentially install each required package using the following:
$ git clone --recursive https://github.com/RaduAlexandru/easy_pbr
$ cd easy_pbr && make && cd ..
$ git clone --recursive https://github.com/RaduAlexandru/data_loaders
$ cd data_loaders && make && cd ..
$ git clone --recursive https://github.com/RaduAlexandru/neural_mvs
$ cd neural_mvs && make && cd ..NeuralMVS uses RGB images together with camera pose for training. The data is loaded with the DataLoaders package and interfaced using EasyPBR. Here we show how to train on the DTU, LLFF and NeRF synthetic dataset.
Download either dataset by clicking on their respective links.
Next modify the neural_mvs/config/train.cfg and for each respective loader modify the dataset_path to the folder in which your dataset is located.
NeuralMVS uses config files to configure the dataset used, the training parameters, model architecture and various visualization options.
The config file used to train can be found under "neural_mvs/config/train.cfg".
Running the training script will by default read this config file and start the training.
$ cd neural_mvs
$ ./neural_mvs_py/train.pyBy default the data is visualized using tensorboard. Therefore we recommend to run the following lines before starting the training script.
$ cd neural_mvs
$ mkdir tensorboard_logs
$ tensorboard --logdir=tensorboard_logs --port=6006The training can be visualized in the browser under http://localhost:6006/
Various configuration options can be interesting to check out and modify. We take neural_mvs/config/train.cfg as an example.
dataset_name: "nerf" #can be changed to dtu or llff to switch the dataset we use
core: hdpi: false #can be turned on an off to accomodate high DPI displays. If the text and fonts in the visualizer are too big, set this option to false
train: with_viewer: false #setting to true will start a visualizer which displays the currently predicted 3D point cloud and image. May not be usable when running on a headless server
@inproceedings{rosu2021neuralmvs,
title={NeuralMVS: Bridging Multi-View Stereo and Novel View Synthesis},
author={Rosu, Radu Alexandru and Behnke, Sven},
booktitle="International Joint Conference on Neural Networks (IJCNN)",
year={2022}
}
NeuralMVS is provided under the terms of the MIT license (see LICENSE)