Skip to content

Code for "Practical Deep Stereo (PDS): Toward applications-friendly deep stereo matching", NeurIPS 2018.

License

Notifications You must be signed in to change notification settings

tlkvstepan/PracticalDeepStereo_NIPS2018

Repository files navigation

Practical Deep Stereo

This repository contains refactored code for "Practical Deep Stereo (PDS): Toward applications-friendly deep stereo matching" by Stepan Tulyakov, Anton Ivanov and Francois Fleuret, that appeared on NeurIPS2018 as a poster.

Requirements

Please install conda with python 3.6 and pytorch 1.0. Next install all dependencies by running

conda install --yes --file requirements.txt

Preparing Datasets

To set up FlyingThings3D dataset, download PNG RGB cleanpass images and disparities from website of Patter Recognition and Image Processing group of University of Freiburg. Unpack the archive with images into PracticalDeepStereo_NIPS2018/datasets/flyingthings3d/frames_cleanpass and archive with disparities into PracticalDeepStereo_NIPS2018/datasets/flyingthings3d/disparity.

Training and benchmarking on FlyingThings3D

To run training on Flyingthings3D run

./train_on_flyingthings3d.py \
--experiment_folder experiments/flyingthings3d \
--dataset_folder datasets/flyingthings3d \

During the first run, the dataset object calculates and saves disparity statistic for every example in the dataset. Therefore, it might take a while before actual training starts. Overall, the training on full-size image is very slow since it does not use batch processing. The training can be started from a checkpoint by setting the --checkpoint_file flag.

To benchmark on Flyingthings3D run

./benchmark_on_flyingthings3d.py \
--experiment_folder experiments/flyingthings3d \
--dataset_folder datasets/flyingthings3d \
--checkpoint_file experiments/flyingthings3d/010_checkpoint.bin \
--is_psm_protocol

The evalutaion protocol can be selected by setting / unsetting the --is_psm_protocol flag.

Pretrained model with training plot and log are now avaliable.

For the pre-trained model test results are following

Protocol MAE, [pix] 3PE, [%] time-per-image, [sec]
PSM 1.04 3.09 0.62
CRL 0.98 2.96

Note, that this result is slightly better than the one reported in the paper according to the PSM protocol and slightly worse according to CRL (but both results are still very good!).

Troubleshooting

If one of the training scripts does not work please run all unit tests by executing ./run_unit_tests.sh. This will help you to localize and fix bugs on your own.

About

Code for "Practical Deep Stereo (PDS): Toward applications-friendly deep stereo matching", NeurIPS 2018.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages