Followed from https://github.com/JiaRenChang/PSMNet

Dependencies

• Python 3.5

• PyTorch 0.4.0+

• torchvision 0.2.0

Dataset

• KITTI 2012 / KITTI 2015

• Scene Flow

KITTI: The KITTI dataset is relatively small and has real-world pictures with sparse ground truth disparity maps.

Scene Flow: The Scene Flow dataset is relatively large. It is a synthetic dataset with dense ground truth disparity maps.

Training Strategy: Pretrain the network using Scene Flow dataset, then finetune the network on the KITTI dataset

Train

python main.py --maxdisp 192 \
               --datapath /SceneFlow_dataset_path \
               --epochs 10  \
               --savemodel ./trained/

python finetune.py --maxdisp 192 \
                   --datatype 2015 \
                   --datapath /KITTI2_train_set_path \
                   --epochs 600 \
                   --loadmodel ./trained/checkpoint_10.tar \
                   --savemodel ./trained/trained2015

Evaluation

python submission.py --loadmodel ./trained/trained2012/finetune_600.tar  \
                     --KITTI 2012
                     --datapath  /KITTI_test_set_path

Final Result

The disparity maps for the KITTI testset are calculated by submission.py, those disparity maps are submitted to the KITTI website to calculated the final accuracy result.

Please cite

@article{luan2018gabor,
  title={Gabor convolutional networks},
  author={Luan, Shangzhen and Chen, Chen and Zhang, Baochang and Han, Jungong and Liu, Jianzhuang},
  journal={IEEE Transactions on Image Processing},
  volume={27},
  number={9},
  pages={4357--4366},
  year={2018},
  publisher={IEEE}
}

@inproceedings{liu2018stereo,
  title={Stereo Matching Using Gabor Convolutional Neural Network},
  author={Liu, Zhendong and Hu, Qinglei and Liu, Jiachen and Zhang, Baochang},
  booktitle={2018 11th International Workshop on Human Friendly Robotics (HFR)},
  pages={48--53},
  year={2018},
  organization={IEEE}
}