NeuralBlox: Real-Time Neural Representation Fusion for Robust Volumetric Mapping

This repository contains NeuralBlox, our framework for volumetric mapping in latent neural representation space.

Table of Contents

• Paper
• Installation
• Demo

Paper

If you find our code or paper useful, please consider citing us:

• Stefan Lionar*, Lukas Schmid*, Cesar Cadena, Roland Siegwart, and Andrei Cramariuc. "NeuralBlox: Real-Time Neural Representation Fusion for Robust Volumetric Mapping", International Conference on 3D Vision (3DV), pp. 1279-1289, 2021. (* equal contribution) [ IEEE | ArXiv | Supplementary ]

  @inproceedings{lionar2021neuralblox,
   title = {NeuralBlox: Real-Time Neural Representation Fusion for Robust Volumetric Mapping},
   author={Stefan Lionar, Lukas Schmid, Cesar Cadena, Roland Siegwart, Andrei Cramariuc},
   booktitle={2021 International Conference on 3D Vision (3DV)}, 
   year={2021},
   pages={1279-1289},
   doi={10.1109/3DV53792.2021.00135}}
  }

Installation

conda env create -f environment.yaml
conda activate neuralblox
pip install torch-scatter==2.0.4 -f https://pytorch-geometric.com/whl/torch-1.4.0+cu101.html

Note: Make sure torch-scatter and PyTorch have the same cuda toolkit version. If PyTorch has a different cuda toolkit version, run:

conda install pytorch==1.4.0 cudatoolkit=10.1 -c pytorch

Next, compile the extension modules. You can do this via

python setup.py build_ext --inplace

Optional: For a noticeably faster inference on CPU-only settings, upgrade PyTorch and PyTorch Scatter to a newer version:

pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 -f https://download.pytorch.org/whl/torch_stable.html
pip install --upgrade --no-deps --force-reinstall torch-scatter==2.0.5 -f https://pytorch-geometric.com/whl/torch-1.7.1+cu101.html

Demo

To generate meshes using our pretrained models and evaluation dataset, you can select several configurations below and run it.

python generate_sequential.py configs/fusion/pretrained/redwood_0.5voxel_demo.yaml
python generate_sequential.py configs/fusion/pretrained/redwood_1voxel_demo.yaml
python generate_sequential.py configs/fusion/pretrained/redwood_1voxel_demo_cpu.yaml --no_cuda

• The mesh will be generated to out_mesh/mesh folder.
• To add noise, change the values under test.scene.noise in the config files.

Training backbone encoder and decoder

The backbone encoder and decoder mainly follow Convolutional Occupancy Networks (https://github.com/autonomousvision/convolutional_occupancy_networks) with some modifications adapted for our use case. Our pretrained model is provided in this repository.

Dataset

ShapeNet

The proprocessed ShapeNet dataset is from Occupancy Networks (https://github.com/autonomousvision/occupancy_networks). You can download it (73.4 GB) by running:

bash scripts/download_shapenet_pc.sh

After that, you should have the dataset in data/ShapeNet folder.

Training

To train the backbone network from scratch, run

python train_backbone.py configs/pointcloud/shapenet_grid24_pe.yaml

Latent code fusion

The pretrained fusion network is also provided in this repository.

Training dataset

To train from scratch, you can download our preprocessed Redwood Indoor RGBD Scan dataset by running:

bash scripts/download_redwood_preprocessed.sh

We align the gravity direction to be the same as ShapeNet ([0,1,0]) and convert the RGBD scans following ShapeNet format.

More information about the dataset is provided here: http://redwood-data.org/indoor_lidar_rgbd/.

Training

To train the fusion network from scratch, run

python train_fusion.py configs/fusion/train_fusion_redwood.yaml

Adjust the path to the encoder-decoder model in training.backbone_file of the .yaml file if necessary.

Generation

python generate_sequential.py CONFIG.yaml

If you are interested in generating the meshes from other dataset, e.g., ScanNet:

• Structure the dataset following the format in demo/redwood_apartment_13k.
• Adjust path, data_preprocessed_interval and intrinsics in the config file.
• If necessary, align the dataset to have the same gravity direction as ShapeNet by adjusting align in the config file.

For example,

# ScanNet scene ID 0
python generate_sequential.py configs/fusion/pretrained/scannet_000.yaml

# ScanNet scene ID 24
python generate_sequential.py configs/fusion/pretrained/scannet_024.yaml

To use your own models, replace test.model_file (encoder-decoder) and test.merging_model_file (fusion network) in the config file to the path of your models.

Evaluation

You can evaluate the predicted meshes with respect to a ground truth mesh by following the steps below:

1. Install CloudCompare

sudo apt install cloudcompare

2. Copy a ground truth mesh (no RGB information expected) to evaluation/mesh_gt
3. Copy prediction meshes to evaluation/mesh_pred
4. If the prediction mesh does not contain RGB information, such as the output from our method, run:

python evaluate.py

Else, if it contains RGB information, such as the output from Voxblox, run:

python evaluate.py --color_mesh

We provide the trimmed mesh used for the ground truth of our quantitative evaluation. It can be downloaded here.

Lastly, to evaluate prediction meshes with respect to the trimmed mesh as ground truth, run:

python evaluate.py --demo

Or for colored mesh (e.g. from Voxblox):

python evaluate.py --demo --color_mesh

evaluation.csv will be generated to evaluation directory.

Acknowledgement

Some parts of the code are inherited from the official repository of Convolutional Occupancy Networks (https://github.com/autonomousvision/convolutional_occupancy_networks).