Stereo Matching Network

Original article: https://arxiv.org/abs/1510.05970

Setup

Requirements:

• PyTorch
• H5Py (with parallelism):

Run

Create HDF5 dataset:

  python create_hdf5.py path/to/kitti/training --dest=path/to/kitti.h5

Train the model:

  python train.py path/to/kitti/training store/dir --train_data=path/to/kitti.h5

TODO

HDF5 reading is slow. Data reader(and filesystem organisation) should be implemented to enable efficient data preprocessing. Currently, whole dataset is loaded in memory to enable efficient GPu utilization(dataset is ~13GB).