Awesome

EASNet

EASNet: Searching Elastic and Accurate Network Architecture for Stereo Matching (ECCV 2022) [arXiv] [Video]

PyTorch implementation of searching an efficient network architecture for stereo matching. Please cite the paper below if you use this project. Any suggestion, fork, and pull request is welcome.

@inproceedings{
  wang2022easnet,
  title={EASNet: Searching Elastic and Accurate Network Architecture for Stereo Matching},
  author={Qiang Wang and Shaohuai Shi and Kaiyong Zhao and Xiaowen Chu},
  booktitle={European Conference on Computer Vision},
  year={2022},
  url={https://arxiv.org/abs/2207.09796}
}

Setup

Python 3 dependencies:

• PyTorch 1.8+
• OpenMPI 4.0.1+
• matplotlib
• numpy
• imageio
• other necessary packages

We use the deformation module from AANet. Install the "deform_conv" package as follow.

cd ofa/stereo_matching/networks/deform_conv
sh build.sh

Our training scripts apply MPI to accelerate the training procedure. Please install OpenMPI 4.0.1 or above.

Searching EASNet

The main commands are summarized in "train.sh". One can use them accordingly.

Train the largest supernet.

# two nodes, four GPUs per node
mpirun -np 8 -H host1:4,host2:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    python train_ofa_stereo.py \
           --task large

Shrink the kernel size/depth/width/scale.

# two nodes, four GPUs per node
export TASK=kernel  # 'kernel', 'depth', 'width', 'scale'
mpirun -np 8 -H host1:4,host2:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    python train_ofa_stereo.py \
           --task $TASK

Finetuning on KITTI 2012/2015

# two nodes, four GPUs per node
export TASK=kitti12  # 'kitti12', 'kitti2015'
mpirun -np 8 -H host1:4,host2:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    python train_ofa_stereo.py \
           --task $TASK

Pretrained Models