This paper has been accepted by ECCV 2022

By Ziyue Feng, Liang Yang, Longlong Jing, Haiyan Wang, Yingli Tian, and Bing Li.

Arxiv: Link Youtube: link Project Site: link

You can install the dependencies with:

conda create -n dynamicdepth python=3.6.6
conda activate dynamicdepth
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge
pip install tensorboardX==1.4
conda install opencv=3.3.1   # just needed for evaluation
pip install open3d
pip install wandb
pip install scikit-image
python -m pip install cityscapesscripts

We ran our experiments with PyTorch 1.8.0, CUDA 11.1, Python 3.6.6 and Ubuntu 18.04.

Pull the repository and make a folder named CS_RAW for cityscapes raw data:

git clone https://github.com/AutoAILab/DynamicDepth.git
cd DynamicDepth
cd data
mkdir CS_RAW

From Cityscapes official website download the following packages: 1) leftImg8bit_sequence_trainvaltest.zip, 2) camera_trainvaltest.zip into the CS_RAW folder.

Preprocess the Cityscapes dataset using the prepare_train_data.py(from SfMLearner) script with following command:

cd CS_RAW
unzip leftImg8bit_sequence_trainvaltest.zip
unzip camera_trainvaltest.zip
cd ..

python prepare_train_data.py \
    --img_height 512 \
    --img_width 1024 \
    --dataset_dir CS_RAW \
    --dataset_name cityscapes \
    --dump_root CS \
    --seq_length 3 \
    --num_threads 8

Download cityscapes depth ground truth(provided by manydepth) for evaluation:

cd ..
cd splits/cityscapes/
wget https://storage.googleapis.com/niantic-lon-static/research/manydepth/gt_depths_cityscapes.zip
unzip gt_depths_cityscapes.zip
cd ../..

(Recommended)Download Manydepth pretrained model from Here and put in the log folder. Training from these weights will converge much faster.

mkdir log
cd log
# Download CityScapes_MR.zip to here 
unzip CityScapes_MR.zip
cd ..

Download dynamic object masks for Cityscapes dataset from (Google Drive or OneDrive) and extract the train_mask and val_mask folder to DynamicDepth/data/CS/.

By default models and log event files are saved to log/dynamicdepth/models.

python -m dynamicdepth.train  # the configs are defined in options.py

val() function in the trainer.py evaluates the model on Cityscapes testing set.

You can download our pretrained model from the following links:

@article{feng2022disentangling,
  title={Disentangling Object Motion and Occlusion for Unsupervised Multi-frame Monocular Depth},
  author={Feng, Ziyue and Yang, Liang and Jing, Longlong and Wang, Haiyan and Tian, YingLi and Li, Bing},
  journal={arXiv preprint arXiv:2203.15174},
  year={2022}
}

InstaDM: https://github.com/SeokjuLee/Insta-DM ManyDepth: https://github.com/nianticlabs/manydepth

If you have any concern with this paper or implementation, welcome to open an issue or email me at '[email protected]'