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: https://bitbucket.org/saakuraa/cvpr16_stereo_public/overview

Architecture of the win37_dep9 network:

Global view

Detailed view

Installation

1. Install Tensorflow

2. Clone the repository

git clone https://github.com/fjulca-aguilar/dl_stereo_matching.git

Training new models

New models can be trained using

python train.py --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 https://bitbucket.org/saakuraa/cvpr16_stereo_public/overview

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

Testing on images

Evaluation on images can be done using

python test_images.py \ 
--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