DRO: Deep Recurrent Optimizer for Structure-from-Motion

This is the official PyTorch implementation code for DRO-sfm. For technical details, please refer to:

DRO: Deep Recurrent Optimizer for Structure-from-Motion
Xiaodong Gu*, Weihao Yuan*, Zuozhuo Dai, Chengzhou Tang, Siyu Zhu, Ping Tan
[Paper]

Bibtex

If you find this code useful in your research, please cite:

@article{gu2021dro,
  title={DRO: Deep Recurrent Optimizer for Structure-from-Motion},
  author={Gu, Xiaodong and Yuan, Weihao and Dai, Zuozhuo and Tang, Chengzhou and Zhu, Siyu and Tan, Ping},
  journal={arXiv preprint arXiv:2103.13201},
  year={2021}
}

Contents

1. Install
2. Datasets
3. Training
4. Evaluation
5. Models

Install

• We recommend using nvidia-docker2 to have a reproducible environment.

git clone https://github.com/aliyun/dro-sfm.git
cd dro-sfm
sudo make docker-build
sudo make docker-start-interactive

You can also download the built docker directly from dro-sfm-image.tar

docker load < dro-sfm-image.tar

• If you do not use docker, you could create an environment following the steps in the Dockerfile.

# Environment variables
export PYTORCH_VERSION=1.4.0
export TORCHVISION_VERSION=0.5.0
export NCCL_VERSION=2.4.8-1+cuda10.1
export HOROVOD_VERSION=65de4c961d1e5ad2828f2f6c4329072834f27661
# Install NCCL
sudo apt-get install libnccl2=${NCCL_VERSION} libnccl-dev=${NCCL_VERSION}

# Install Open MPI
mkdir /tmp/openmpi && \
    cd /tmp/openmpi && \
    wget https://www.open-mpi.org/software/ompi/v4.0/downloads/openmpi-4.0.0.tar.gz && \
    tar zxf openmpi-4.0.0.tar.gz && \
    cd openmpi-4.0.0 && \
    ./configure --enable-orterun-prefix-by-default && \
    make -j $(nproc) all && \
    make install && \
    ldconfig && \
    rm -rf /tmp/openmpi

# Install PyTorch
pip install torch==${PYTORCH_VERSION} torchvision==${TORCHVISION_VERSION} && ldconfig

# Install horovod (for distributed training)
sudo ldconfig /usr/local/cuda/targets/x86_64-linux/lib/stubs && HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir git+https://github.com/horovod/horovod.git@${HOROVOD_VERSION} && sudo ldconfig

To verify that the environment is setup correctly, you can run a simple overfitting test:

# download a tiny subset of KITTI
cd dro-sfm
curl -s https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/dro-sfm/datasets/KITTI_tiny.tar | tar xv -C /data/datasets/kitti/
# in docker
./run.sh "python scripts/train.py configs/overfit_kitti_mf_gt.yaml" log.txt

Datasets

Datasets are assumed to be downloaded in /data/datasets/<dataset-name> (can be a symbolic link).

KITTI

The KITTI (raw) dataset used in our experiments can be downloaded from the KITTI website. For convenience, you can download data from packnet or here

Tiny KITTI

For simple tests, you can download a "tiny" version of KITTI:

Scannet

The Scannet (raw) dataset used in our experiments can be downloaded from the Scannet website. For convenience, you can download data from here

DeMoN

Download DeMoN.

bash download_traindata.sh
python ./dataset/preparation/preparedata_train.py
bash download_testdata.sh
python ./dataset/preparation/preparedata_test.py

Training

Any training, including fine-tuning, can be done by passing either a .yaml config file or a .ckpt model checkpoint to scripts/train.py:

# kitti, checkpoints will saved in ./results/mdoel/
./run.sh 'python scripts/train.py  configs/train_kitti_mf_gt.yaml' logs/kitti_sup.txt
./run.sh 'python scripts/train.py  configs/train_kitti_mf_selfsup.yaml' logs/kitti_selfsup.txt 

# scannet
./run.sh 'python scripts/train.py  configs/train_scannet_mf_gt_view3.yaml' logs/scannet_sup.txt
./run.sh 'python scripts/train.py  configs/train_scannet_mf_selfsup_view3.yaml' logs/scannet_selfsup.txt
./run.sh 'python scripts/train.py  configs/train_scannet_mf_gt_view5.yaml' logs/scannet_sup_view5.txt

# demon
./run.sh 'python scripts/train.py  configs/train_demon_mf_gt.yaml' logs/demon_sup.txt

Evaluation

python scripts/eval.py --checkpoint <checkpoint.ckpt> [--config <config.yaml>]
# example:kitti, results will be saved in results/depth/
python scripts/eval.py --checkpoint ckpt/outdoor_kitti.ckpt --config configs/train_kitti_mf_gt.yaml

You can also directly run inference on a single image or video:

# video or folder
# indoor-scannet 
python scripts/infer_video.py --checkpoint ckpt/indoor_sacnnet.ckpt --input /path/to/video or folder --output /path/to/save_folder --sample_rate 1 --data_type scannet --ply_mode 
 # indoor-general
python scripts/infer_video.py --checkpoint ckpt/indoor_sacnnet.ckpt --input /path/to/video or folder --output /path/to/save_folder --sample_rate 1 --data_type general --ply_mode

# outdoor
python scripts/infer_video.py --checkpoint ckpt/outdoor_kitti.ckpt --input /path/to/video or folder --output /path/to/save_folder --sample_rate 1 --data_type kitti --ply_mode 

# image
python scripts/infer.py --checkpoint <checkpoint.ckpt> --input <image or folder> --output <image or folder>

Models

Model	Abs.Rel.	Sqr.Rel	RMSE	RMSElog	a1	a2	a3	SILog	L1_inv	rot_ang	t_ang	t_cm
Kitti_sup	0.045	0.193	2.570	0.080	0.971	0.994	0.998	0.079	0.003	-	-	-
Kitti_selfsup	0.053	0.346	3.037	0.102	0.962	0.990	0.996	0.101	0.004	-	-	-
scannet_sup	0.053	0.017	0.165	0.080	0.967	0.994	0.998	0.078	0.033	0.472	9.297	1.160
scannet_sup(view5)	0.047	0.014	0.151	0.072	0.976	0.996	0.999	0.071	0.030	0.456	8.502	1.163
scannet_selfsup	0.143	0.345	0.656	0.274	0.896	0.954	0.969	0.272	0.106	0.609	10.779	1.393

Acknowledgements

Thanks to Toyota Research Institute for opening source of excellent work packnet-sfm. Thanks to Zachary Teed for opening source of his excellent work RAFT.