PermutoSDF: Fast Multi-View Reconstruction with Implicit Surfaces using Permutohedral Lattices
Radu Alexandru Rosu, Sven Behnke
University of Bonn, Autonomous Intelligent Systems

PermutoSDF is a system capable of recovering highly detailed 3D reconstructions given only 2D images with poses. We propose to use a hashed permutohedral encoding to ensure fast training and a novel RGB regularizer to encourage the network to predict high-frequency geometric detail. We model the scene as two neural networks that predict an SDF and a color field. Volume rendering is used to train the two networks end-to-end. After training we can switch to sphere tracing and render at interactive speeds.

Since PermutoSDF requires PyTorch to be installed with CXX_ABI=1, it's best to use the provided dockerfile:

$ git clone --recursive https://github.com/RaduAlexandru/permuto_sdf
$ cd permuto_sdf/docker
$ ./build.sh
$ ./run.sh

This will build a docker image containing almost everything that is needed.

PermutoSDF also depends on EasyPBR, DataLoaders and permutohedral_encoding packages. Once you are inside the docker container (after running ./docker/run.sh) they can be easily installed with the following lines:

$ 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/permutohedral_encoding
$ cd permutohedral_encoding && make && cd ..

We use EasyPBR for visualizing locally the training progress, DataLoaders for loading various datasets and permutohedral_encoding as a self-contained package for performing hash-based encoding using the permutohedral lattice.

After they are installed, one can install permuto_sdf package using

$ git clone --recursive https://github.com/RaduAlexandru/permuto_sdf
$ cd permuto_sdf && make 

Optionally you can also install APEX for slightly faster training. PermutoSDF will automatically detect that it is installed and used their FusedAdam for optimizing.

For training and experiments we use the the Multiface, DTU and BlendedMVS datasets.
We also provide our own EasyPBR-based dataset containing the synthetic head.
All the data used in the paper in a format that is readable by our scripts can be found here:

Unzip all the files datasets inside the permuto_sdf_data folder so you end up with a structure like:

    permuto_sdf_data
    ├── data_BlendedMVS
    │   ├── bmvs_bear
    │   ├── bmvs_clock
    │   └── ...
    ├── data_DTU
    │   ├── dtu_scan24
    │   ├── dtu_scan37
    │   └── ...
    ├── easy_pbr_renders
    │   ├── head
    ├── horse_gallop

Finally you have to open permuto_sdf_py/paths/data_paths.py and add another line for where to find the path permuto_sdf_data on your computer. So the new line will be something of the form comp_3": "<THE_PATH_WHERE_YOU_DOWNLOADED_THE_DATA>/permuto_sdf_data", Now when calling the training routines from the next sections you can add the flag --comp_name comp_3 do indicate that you are going to use this path

./permuto_sdf_py/train_permuto_sdf.py \
--dataset dtu \
--scene dtu_scan24 \
--comp_name comp_3 \
--exp_info default

The training will start and on the dtu_scan24 scene and a viewer will appear where you can inspect the training progress. Other options for this script are:

--no_viewer  #disables the OpenGL viewer so you can run the script on a headless machine
--low_res    #loads the images at lower resolution. Useful on machines with lower than 6GB of VRAM
--with_mask  #Uses the binary mask from DTU or BlendedMVS dataset to ignore the background

Additionally you can enable saving of the checkpoints and other logging options by setting to true the flag save_checkpoint in ./config/train_permuto_sdf.cfg.
At any point during training you can click inside the GUI on "control_view" in order to move the camera around the scene.

./permuto_sdf_py/train_sdf_from_mesh.py \

It will load a high-resolution mesh from EasyPBR and trains only the SDF network.

./permuto_sdf_py/train_4d_sdf.py \
--comp_name comp_3

It will load various meshes from an animation and it will learn the 4D (3D+time) geometry. Using the viewer you can modify the time value so you can sweep through the time-dimension of the model.

./permuto_sdf_py/train_nerf.py \
--dataset dtu \
--scene dtu_scan24 \
--comp_name comp_3 \
--exp_info default

We also provide an example which trains a NeRF-like model using the permutohedral lattice.

It is common to want to train on all the objects from the DTU dataset and instead of running various times the train_permutosdf.py script we provide a script to run through all the object from the dataset one-by-one so you can leave this script running overnight.

./permuto_sdf_py/experiments/serial_training/serial_train.py \
--dataset dtu \
--with_mask \
--comp_name comp_3 \ 
--exp_info experiment_name

We gather the resulting novel-view synthesis images and meshes generated by PermutoSDF under the results folder in here.
We also provide checkpoints for all the objects in the DTU dataset which can be found under the same link.
The results can be regenerated from the checkpoints by using the scripts in create_my_images.py and create_my_meshes.py from the folder ./permuto_sdf_py/experiments/evaluation. Afterwards the evaluate_psnr.py and evaluate_chamfer_distance.py should report the corresponding metrics.

PermutoSDF is provided under the terms of the MIT license (see LICENSE).

@inproceedings{rosu2023permutosdf,
    title={PermutoSDF: Fast Multi-View Reconstruction with 
            Implicit Surfaces using Permutohedral Lattices  },
    author={Radu Alexandru Rosu and Sven Behnke},
    booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2023}
}