Stereo Plane R-CNN

Paper

If you find Stereo Plane R-CNN useful in your academic work please cite the following paper:

@article{wietrzykowski2022,
    author={Wietrzykowski, Jan and Belter, Dominik},
    journal={IEEE Robotics and Automation Letters}, 
    title={{Stereo Plane R-CNN: Accurate Scene Geometry Reconstruction Using Planar Segments and Camera-Agnostic Representation}}, 
    year={2022},
    volume={7},
    number={2},
    pages={4345-4352},
    doi={10.1109/LRA.2022.3150841}
 }

Instalation

Clone the repository:

git clone https://github.com/LRMPUT/sprcnn.git

Create a new Conda environment and install required dependencies:

conda install pytorch torchvision torchaudio torchmetrics cudatoolkit=11.1 numpy scikit-image scipy numpy -c pytorch-lts -c nvidia

Install other depencencies using Pip:

pip install opencv-python disjoint-set pytorch-lightning

Dataset

The SceneNet Stereo dataset can be downloaded here.

The TERRINet dataset can be downloaded here.

Training

python train_planercnn.py --dataFolder=/mnt/data/datasets/scenenet_rgbd --anchorType=none_exp_plane_params --normWeight=100.0 --dispWeight=1.0 --LR=0.00001 --numEpochs=10

Inference

Pretrained model can be downloaded here.

python evaluate.py --anchorType=none_exp_plane_params --dataFolder=/mnt/data/datasets/TERRINet --checkpoint=/mnt/data/datasets/scenenet_rgbd/checkpoint/plane_params.ckpt --no_normals

Exporting detections for PlaneLoc2

Before running the following command, make sure that contents of annotations_plane_params_det dir in each scene directory is empty.

python evaluate.py --anchorType=none_exp_plane_params --dataFolder=/mnt/data/datasets/TERRINet --checkpoint=/mnt/data/datasets/scenenet_rgbd/checkpoint/plane_params.ckpt --no_normals --export_detections

Detections should be exported to annotations_plane_params_det dir in each scene directory.

Acknowledgement

Our implementation is based on Plane R-CNN.