DROID-SLAM

DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras
Zachary Teed and Jia Deng

@article{teed2021droid,
  title={{DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras}},
  author={Teed, Zachary and Deng, Jia},
  journal={Advances in neural information processing systems},
  year={2021}
}

Initial Code Release: This repo currently provides a single GPU implementation of our monocular, stereo, and RGB-D SLAM systems. It currently contains demos, training, and evaluation scripts.

Requirements

To run the code you will need ...

Inference: Running the demos will require a GPU with at least 11G of memory.
Training: Training requires a GPU with at least 24G of memory. We train on 4 x RTX-3090 GPUs.

Getting Started

Clone the repo using the --recursive flag

git clone --recursive https://github.com/princeton-vl/DROID-SLAM.git

Creating a new anaconda environment using the provided .yaml file. Use environment_novis.yaml to if you do not want to use the visualization

# updated
conda env create -f environment.yaml
conda activate droidenv
conda install cudatoolkit-dev=11.3
pip install evo --upgrade --no-binary evo
pip install gdown
pip install open3d -U

Compile the extensions (takes about 10 minutes)

python setup.py install

Demos

Download the model from google drive: droid.pth
Download some sample videos using the provided script.

./tools/download_sample_data.sh

Run the demo on any of the samples (all demos can be run on a GPU with 11G of memory). While running, press the "s" key to increase the filtering threshold (= more points) and "a" to decrease the filtering threshold (= fewer points). To save the reconstruction with full resolution depth maps use the --reconstruction_path flag.

python demo.py --imagedir=data/abandonedfactory --calib=calib/tartan.txt --stride=2

python demo.py --imagedir=data/sfm_bench/rgb --calib=calib/eth.txt

python demo.py --imagedir=data/Barn --calib=calib/barn.txt --stride=1 --backend_nms=4

python demo.py --imagedir=data/mav0/cam0/data --calib=calib/euroc.txt --t0=150

python demo.py --imagedir=data/rgbd_dataset_freiburg3_cabinet/rgb --calib=calib/tum3.txt

Running on your own data: All you need is a calibration file. Calibration files are in the form

fx fy cx cy [k1 k2 p1 p2 [ k3 [ k4 k5 k6 ]]]

with parameters in brackets optional.

Evaluation

We provide evaluation scripts for TartanAir, EuRoC, and TUM. EuRoC and TUM can be run on a 1080Ti. The TartanAir and ETH will require 24G of memory.

TartanAir (Mono + Stereo)

Download the TartanAir dataset using the script thirdparty/tartanair_tools/download_training.py and put them in datasets/TartanAir

./tools/validate_tartanair.sh --plot_curve            # monocular eval
./tools/validate_tartanair.sh --plot_curve  --stereo  # stereo eval

EuRoC (Mono + Stereo)

Download the EuRoC sequences (ASL format) and put them in datasets/EuRoC

./tools/evaluate_euroc.sh                             # monocular eval
./tools/evaluate_euroc.sh --stereo                    # stereo eval

TUM-RGBD (Mono)

Download the fr1 sequences from TUM-RGBD and put them in datasets/TUM-RGBD

./tools/evaluate_tum.sh                               # monocular eval

FusionPortable (Mono)

Download the original ros-bag from FusionPortableV2

Extract images from original ros-bag

python tools/bag_to_images.py --bag_file=xx/ --output_dir=xx/ --topic=/stereo/frame_left/image_raw/compressed --timestamp=xx/

Run Droid-SLAM

python evaluation_scripts/test_fp.py --root=/Path/to/fusionportable/data --weights=droid.pth --calib=calib/fusionportable_left.txt --rgb_list=frame_left.txt

ETH3D (RGB-D)

Download the ETH3D dataset

./tools/evaluate_eth3d.sh                             # RGB-D eval

Training

First download the TartanAir dataset. The download script can be found in thirdparty/tartanair_tools/download_training.py. You will only need the rgb and depth data.

python download_training.py --rgb --depth

You can then run the training script. We use 4x3090 RTX GPUs for training which takes approximatly 1 week. If you use a different number of GPUs, adjust the learning rate accordingly.

Note: On the first training run, covisibility is computed between all pairs of frames. This can take several hours, but the results are cached so that future training runs will start immediately.

python train.py --datapath=<path to tartanair> --gpus=4 --lr=0.00025

Acknowledgements

Data from TartanAir was used to train our model. We additionally use evaluation tools from evo and tartanair_tools.

jarvisyjw/DROID-SLAM