tonyjo / edlsm_pytorch Goto Github PK

This is a Pytorch implementation for stereo matching described in the paper: Efficient Deep learning for stereo matching

1. Pytorch Version --v0.4 with CUDA > 8.0
2. Numpy --v1.14
3. OpenCV --v3.2
4. Matplotlib --v2.1

First we need to generate the valid pixel locations.

1. Download the KITTI 2015 dataset.
2. Go to the cloned EDLSM folder and run the following command:

mkdir dataset

python prepare_kitti_dataset.py  --dataset_dir=/path/to/data_scene_flow/training --dump_root=./dataset

Note: In order to see other changeable parameters such as patch size, image height/width, train/val split etc, run the following command:

python prepare_kitti_dataset.py  --h

Once the data is successfully prepared, the model can be trained by running the following command:

python train.py --dataset_dir=/path/to/data_scene_flow/training --train_dataset_name=tr_160_18_100.txt --checkpoint_dir=/where/to/store/checkpoints

Note: In order to see other changeable parameters such as batch size, image height/width, pixel weights, etc., run the following command:

python prepare_kitti_dataset.py  --h

In order to see the training loss graph open a tensorboard session by

tensorboard --logdir=./logs --port 8080

Once model is trained, we can generate disparity map by running the following command:

python inference_match.py --dataset_dir=/path/to/data_scene_flow/training --checkpoint_dir=/where/checkpoints/stored --checkpoint_file=eldsm_38000 --test_num=82

Note: In order to see other changeable parameters such as image height/width,disparity range, etc, run the following command:

python inference_match.py  --h

Implement the post-processing steps.

Original Code: https://bitbucket.org/saakuraa/cvpr16_stereo_public

Pytorch Code: https://github.com/vijaykumar01/stereo_matching

@inproceedings{luo2016efficient,
  title={Efficient deep learning for stereo matching},
  author={Luo, Wenjie and Schwing, Alexander G and Urtasun, Raquel},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={5695--5703},
  year={2016}
}