Deep stereo matching using Tensorflow

This is a TensorFlow implementation of the stereo matching algorithm described in the paper "Efficient Deep Learning for Stereo Matching".

Compatibility

The code is tested using Tensorflow r1.4 under Ubuntu 14.04 with Python 2.7.

Training data

The KITTI 2015 dataset has been used for training. This dataset consists of total of 200 scenes for training and of 200 scenes for testing. For more details, please check the KITTI website.

Pre-processing

For training and validation, locations from the ground truth disparity images are generated using the preprocessing scripts published in the paper code available here. This preprocessing script generates 3 binary files. Only the training and validation binary files are used.

Running training

You can run the training as follows :

python train_sm.py --data_dir data_path \
--log_dir log_dir \
--train_loc train_binary_file\
--valid_loc validation_binary_file &

You can skip validation by removing the "valid_loc" argument.

Performance on test images

To generate stereo estimates using the pretrained model, this can be done using

python validate_on_test_images.py --data_dir test_data_path \
--model mode_dir \

This script will generate stereo estimates for the whole testing images set.