Meta AI Research, FAIR; University of Oxford, VGG
Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, Christian Rupprecht
DynamicStereo is a transformer-based architecture for temporally consistent depth estimation from stereo videos. It has been trained on a combination of two datasets: SceneFlow and Dynamic Replica that we present below.
dynamic_replica.mp4
The dataset consists of 145200 stereo frames (524 videos) with humans and animals in motion.
We provide annotations for both left and right views, see this notebook:
- camera intrinsics and extrinsics
- image depth (can be converted to disparity with intrinsics)
- instance segmentation masks
- binary foreground / background segmentation masks
- optical flow (released!)
- long-range pixel trajectories (released!)
Download links.json from the data tab on the project website after accepting the license agreement.
git clone https://github.com/facebookresearch/dynamic_stereo
cd dynamic_stereo
export PYTHONPATH=`(cd ../ && pwd)`:`pwd`:$PYTHONPATH
Add the downloaded links.json file to the project folder. Use flag download_splits to choose dataset splits that you want to download:
python ./scripts/download_dynamic_replica.py --link_list_file links.json \
--download_folder ./dynamic_replica_data --download_splits real valid test train
Memory requirements for dataset splits after unpacking (with all the annotations):
- train - 1.8T
- test - 328G
- valid - 106G
- real - 152M
You can use this PyTorch dataset class to iterate over the dataset.
Describes installation of DynamicStereo with the latest PyTorch3D, PyTorch 1.12.1 & cuda 11.3
git clone https://github.com/facebookresearch/dynamic_stereo
cd dynamic_stereo
export PYTHONPATH=`(cd ../ && pwd)`:`pwd`:$PYTHONPATH
conda create -n dynamicstereo python=3.8
conda activate dynamicstereo
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
# It will require some time to install PyTorch3D. In the meantime, you may want to take a break and enjoy a cup of coffee.
conda install -c bottler nvidiacub
python -m pip install "git+https://github.com/facebookresearch/[email protected]"
python -m pip install -r requirements.txt
mkdir third_party
cd third_party
git clone https://github.com/princeton-vl/RAFT-Stereo
cd RAFT-Stereo
bash download_models.sh
cd ../..
To download the checkpoints, you can follow the below instructions:
mkdir checkpoints
cd checkpoints
wget https://dl.fbaipublicfiles.com/dynamic_replica_v1/dynamic_stereo_sf.pth
wget https://dl.fbaipublicfiles.com/dynamic_replica_v1/dynamic_stereo_dr_sf.pth
cd ..
You can also download the checkpoints manually by clicking the links below. Copy the checkpoints to ./dynamic_stereo/checkpoints.
- DynamicStereo trained on SceneFlow
- DynamicStereo trained on SceneFlow and Dynamic Replica
To evaluate DynamicStereo:
python ./evaluation/evaluate.py --config-name eval_dynamic_replica_40_frames \
MODEL.model_name=DynamicStereoModel exp_dir=./outputs/test_dynamic_replica_ds \
MODEL.DynamicStereoModel.model_weights=./checkpoints/dynamic_stereo_sf.pth
Due to the high image resolution, evaluation on Dynamic Replica requires a 32GB GPU. If you don't have enough GPU memory, you can decrease kernel_size from 20 to 10 by adding MODEL.DynamicStereoModel.kernel_size=10 to the above python command. Another option is to decrease the dataset resolution.
As a result, you should see the numbers from Table 5 in the paper. (for this, you need kernel_size=20)
Reconstructions of all the Dynamic Replica splits (including real) will be visualized and saved to exp_dir.
If you installed RAFT-Stereo, you can run:
python ./evaluation/evaluate.py --config-name eval_dynamic_replica_40_frames \
MODEL.model_name=RAFTStereoModel exp_dir=./outputs/test_dynamic_replica_raft
Other public datasets we use:
Training requires a 32GB GPU. You can decrease image_size and / or sample_len if you don't have enough GPU memory.
You need to donwload SceneFlow before training. Alternatively, you can only train on Dynamic Replica.
python train.py --batch_size 1 \
--spatial_scale -0.2 0.4 --image_size 384 512 --saturation_range 0 1.4 --num_steps 200000 \
--ckpt_path dynamicstereo_sf_dr \
--sample_len 5 --lr 0.0003 --train_iters 10 --valid_iters 20 \
--num_workers 28 --save_freq 100 --update_block_3d --different_update_blocks \
--attention_type self_stereo_temporal_update_time_update_space --train_datasets dynamic_replica things monkaa driving
If you want to train on SceneFlow only, remove the flag dynamic_replica from train_datasets.
The majority of dynamic_stereo is licensed under CC-BY-NC, however portions of the project are available under separate license terms: RAFT-Stereo is licensed under the MIT license, LoFTR and CREStereo are licensed under the Apache 2.0 license.
If you use DynamicStereo or Dynamic Replica in your research, please use the following BibTeX entry.
@article{karaev2023dynamicstereo,
title={DynamicStereo: Consistent Dynamic Depth from Stereo Videos},
author={Nikita Karaev and Ignacio Rocco and Benjamin Graham and Natalia Neverova and Andrea Vedaldi and Christian Rupprecht},
journal={CVPR},
year={2023}
}