DROID-SLAM

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}
}

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

If you forgot --recursive

git submodule update --init --recursive .

Installing

Requires CUDA to be installed on your machine. If you run into issues, make sure the PyTorch and CUDA major versions match with the following check (minor version mismatch should be fine).

nvidia-smi
python -c "import torch; print(torch.version.cuda)"

python3 -m venv .venv
source .venv/bin/activate

# install requirements (tested up to torch 2.7)
pip install -r requirements.txt

# optional (for visualization)
pip install moderngl moderngl-window

# install third-party modules (this will take a while)
pip install thirdparty/lietorch
pip install thirdparty/pytorch_scatter

# install droid-backends
pip install -e .

Demos

1. Download the model from google drive: droid.pth or with

   bash复制

   ./tools/download_model.sh

2. Download some sample videos using the provided script.

   bash复制

   ./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). To save the reconstruction with full resolution depth maps use the --reconstruction_path flag. If you ran with --reconstruction_path my_reconstruction.pth, you can view the reconstruction in high resolution by running

python view_reconstruction.py my_reconstruction.pth

Asynchronous and Multi-GPU Inference: You can run the demos in asynchronous mode by running with --asynchronous. In this setting, the frontend and backend will run in seperate Python processes. You can additionally enable multi-GPU inference by setting the devices of the frontend and backend processes with the following arguments.

Visualization currently doesn't work multi-gpu setting. You will need to run with --disable_vis.

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

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, TUM, and ETH3D-SLAM. EuRoC and TUM can be run on a 1080Ti. The TartanAir and ETH3D-SLAM datasets will require 24G of memory.

Asynchronous and Multi-GPU Inference: You can run evaluation in asynchronous mode by running with --asynchronous. In this setting, the frontend and backend will run in seperate Python processes. You can additionally enable multi-GPU inference by setting the devices of the frontend and backend processes with the following arguments. For example:

python evaluation_scripts/test_tartanair.py \
  --datapath data/tartanair_test/mono \
  --gt_path data/tartanair_test/mono_gt \
  --frontend_device cuda:0 \
  --backend_device cuda:1 \
  --asynchronous \
  --disable_vis

Note: Running with --asynchronous will typically produce better results, but this mode is not deterministic.

TartanAir (Mono + Stereo)

Download the TartanAir test set with this command.

./tools/download_tartanair_test.sh

Or from these links: Images, Groundtruth

Monocular evaluation:

python evaluation_scripts/test_tartanair.py \
  --datapath datasets/tartanair_test/mono \
  --gt_path datasets/tartanair_test/mono_gt \
  --disable_vis

Stereo evaluation:

python evaluation_scripts/test_tartanair.py \
  --datapath datasets/tartanair_test/stereo \
  --gt_path datasets/tartanair_test/stereo_gt \
  --stereo --disable_vis

Evaluating on the validation split:

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

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

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

EuRoC (Mono + Stereo)

Download the EuRoC sequences (ASL format):

./tools/download_euroc.sh

Then run evaluation:

# monocular eval (single gpu)
./tools/evaluate_euroc.sh

# monocular eval (multi gpu)
./tools/evaluate_euroc.sh --asynchronous --frontend_device cuda:0 --backend_device cuda:1

# stereo eval (single gpu)
./tools/evaluate_euroc.sh --stereo

# stereo eval (multi gpu)
./tools/evaluate_euroc.sh --stereo --asynchronous --frontend_device cuda:0 --backend_device cuda:1

TUM-RGBD (Mono)

Download the TUM-RGBD sequences:

./tools/download_tum.sh

Then run evaluation:

# monocular eval (single gpu)
./tools/evaluate_tum.sh

# monocular eval (multi gpu)
./tools/evaluate_tum.sh --asynchronous --frontend_device cuda:0 --backend_device cuda:1

ETH3D (RGB-D)

Download the ETH3D dataset:

./tools/download_eth3d.sh

# RGB-D eval (single gpu)
./tools/evaluate_eth3d.sh > eth3d_results.txt
python evaluation_scripts/parse_results.py eth3d_results.txt

# RGB-D eval (multi gpu)
./tools/evaluate_eth3d.sh --asynchronous --frontend_device cuda:0 --backend_device cuda:1 > eth3d_results_async.txt
python evaluation_scripts/parse_results.py eth3d_results_async.txt

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.