Skip to content
A Tensorflow implementation of the models described in the paper "Efficient Deep Learning for Stereo Matching"
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Siamese Deep Neural Networks for Stereo Matching

A Tensorflow implementation of the models described in the paper Efficient Deep Learning for Stereo Matching. This implementation is based on the one provided by the authors of the paper at:

Architecture of the win37_dep9 network:

Global view

Detailed view


  1. Install Tensorflow

  2. Clone the repository

git clone

Training new models

New models can be trained using

python --util_root=PATH_BINARY \
--data_root=PATH_DATABASE \
--net_type=win19_dep9 \
--patch_size=19 \
--model_dir=MODEL_DIR \
--phase=train &

The training and evaluation schemes use the same training data and similar parameters to the ones defined at

Training evolution and graphs can be seen using Tensorboard. The following image shows examples of Cross-entropy Loss evolution for 40 000 training steps (horizontal axes represent the the number of iterations * 100): (blue = win19_dep9, green = win37_dep9)

Evaluation on validation patches

Trained models can be evaluated on validation patches using

python --util_root=PATH_BINARY \
--data_root=PATH_DATABASE \
--net_type=win19_dep9 \
--phase=evaluate &

Testing on images

Evaluation on images can be done using

python \ 
--out_dir=OUT_IMAGES_DIR \
--model_dir=MODEL_DIR \
--data_root=PATH_DATABASE \
--util_root= PATH_BINARY\
--net_type=win19_dep9 \
--patch_size=19 \
--num_imgs=10 &

The script generates num_imgs disparity images and saves them at the OUT_IMAGES_DIR directory

You can’t perform that action at this time.