iSDF

Real-time Neural Signed Distance Fields for Robot Perception

Project page | Paper

iSDF is real-time system the reconstructs the signed distance field (SDF) of room scale environments through the online training of a neural SDF. The model is represented by a MLP that is trained in a continual learning style from a live stream of posed depth images using our self-supervised loss. For details see our paper.

Example results

ReplicaCAD	ScanNet

Two iSDF Modes

There're two options for running iSDF, chose the desired mode and then follow only instructions in that section:

1. Running iSDF on sequences with pose provided: for running iSDF on ScanNet and ReplicaCAD sequences used in the paper.
2. Running iSDF with a live camera in ROS: a ros wrapper for iSDF subscribes to a topic of posed frames that is published by an ORB-SLAM3 wrapper.

1. Running iSDF on sequences with pose provided

Installation

Please read the iSDF modes section to make sure this is the mode you want before installing!

git clone https://github.com/facebookresearch/iSDF.git && cd iSDF

Setup the environment.

conda env create -f environment.yml
conda activate isdf

Install pytorch by following instructions here, then:

pip install -e .

Downloading the data

Make sure to be in the conda environment. To download a single replicaCAD sequence (5GB):

bash data/download_apt_2_nav.sh

To download all 12 sequences (15GB):

bash data/download_data.sh

The data is shared under the same license as the Replica-CAD datset. For instructions on how to generate the dataset see the data README.

ScanNet sequences

To run the ScanNet sequences, you must download the ScanNet dataset. After downloading the scans, you should export the depth, color, poses and intrinsics for the sequences you wish to use with this script.

Usage

If using ScanNet data, then you must set the directory for the downloaded sequence via the key scannet_dir in the config file.

cd isdf/train/
python train_vis.py --config configs/replicaCAD.json

Reproducing experiments

Run a batch of iSDF experiments sequentially in headless mode. To run these experiments you must have downloaded all 12 sequences using our bash script as well as separately downloading and exporting the ScanNet sequences. To run only the ReplicaCAD sequences, you can modify the load_params function in batch_utils.py. If you have multiple GPUs you may want to parallelise the experiment runs. Make sure to update project_dir and scannet_dir in jobs_local.py before running:

cd isdf/train
python batch_train/jobs_local.py

We have provided results for the baselines that we use in the paper for those wanted to reproduce results for iSDF only. Code to reproduce results for our baselines is coming soon.

Below we provide instructions to generate quantitative and qualitative results as in the paper. Note results may vary depending on the CPU / GPU used.

SDF accuracy plots

We provide the script to generate plots comparing SDF error, collision cost error and gradient cosine distance for iSDF and the two baselines. This script was used to generate all quantitative plots in the paper (e.g. Fig 8). Change the variable isdf_dir in the script before running:

python isdf/eval/figs/all_seq.py

2. Running iSDF with a live camera in ROS

We use ROS to interface iSDF with a tracking system and enable running with a live camera. The iSDF node subscribes to the topic: frames, which is a custom message type containing a time synchronised rgb image, depth image and pose. We use a ROS wrapper for ORB-SLAM3 in RGBD mode to publish to this topic. We have developed this project on Ubuntu 20.04 (and ROS Noetic, so several code changes may be needed to adapt to other OS and ROS versions.

Installation

Clone and build ORB-SLAM3

You may need to install some dependencies first, see full instructions if you run into issues.

git clone https://github.com/UZ-SLAMLab/ORB_SLAM3.git ORB_SLAM3
cd ORB_SLAM3
chmod +x build.sh
./build.sh

ROS Noetic

Follow install instructions. Also install catkin tools if not already installed: sudo apt-get install python3-catkin-tools.

Realsense camera

sudo apt-get install ros-$ROS_DISTRO-realsense2-camera

iSDF and ORB-SLAM3 wrapper

Create a catkin workspace and clone iSDF inside:

mkdir -p ~/catkin_ws/src
cd ~/catkin_ws
catkin init
cd src
git clone --recurse-submodules https://github.com/facebookresearch/iSDF.git

Open CMakeLists.txt and change the directory that points to ORB-SLAM3 library at the beginning of the file (default is home folder).

nano ~/catkin_ws/src/iSDF/ORB_SLAM3_ros/CMakeLists.txt

# Change this to your installation of ORB-SLAM3. Default is ~/
set(ORB_SLAM3_DIR
   $ENV{HOME}/ORB_SLAM3
)

Unzip the ORBvoc.txt file:

cd ~/catkin_ws/src/iSDF/ORB_SLAM3_ros/config
tar -xf ORBvoc.txt.tar.gz

Build the ROS packages (while not in a conda environment):

cd ~/catkin_ws
catkin_make
source devel/setup.bash

Setup the iSDF conda environment.

conda env create -f environment.yml
conda activate isdf

Install pytorch by following instructions here, then:

pip install -e .

Usage

Before running you need to modify the camera intrinsics for both iSDF and ORB-SLAM3. If you launch the camera (roslaunch realsense2_camera rs_camera.launch), the intrinsics are published at /camera/color/camera_info. Copy these intrinsics into the iSDF config (isdf/train/configs/realsense.json) and the ORBSLAM3 config (ORB_SLAM3_ros/config/realsense_config.yaml).

To launch run:

roslaunch isdf train.launch show_orbslam_vis:=true

It can be helpful to show the ORB-SLAM3 vis as if the tracking fails iSDF will also fail.

During live operation:

• Press c to clear the keyframe set.
• Press s to pause live operation and open a 3D visualization window. You can change the 3D visualization using key presses, instructions are printed in the terminal when the window is opened.

Our ORB-SLAM3 wrapper also contains a launchfile to run with the Azure Kinect camera. To run with a Kinect camera, modify the train.launch to launch run_kinect.launch and install the kinect ros drivers.

Current limitations:

• iSDF runs in a single thread and alternates between training (optimising the map) and updating the live visualisation. You can change the frequency of the live visualisation updates, however if set too high, training can become slow.
• The live visualistaion of the rendered depth and normals from the current view uses a very basic depth rendering method. This can lead to artifacts in the render even if the geometry is good. We plan to improve the depth rendering soon.

Citation

@inproceedings{Ortiz:etal:iSDF2022,
  title={iSDF: Real-Time Neural Signed Distance Fields for Robot Perception},
  author={Ortiz, Joseph and Clegg, Alexander and Dong, Jing and Sucar, Edgar and Novotny, David and Zollhoefer, Michael and Mukadam, Mustafa},
  booktitle={Robotics: Science and Systems},
  year={2022}
}

License

iSDF is released under the MIT license. Please see the LICENSE file for more information.

Contributing

We actively welcome your pull requests! Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.