This project is to try to recreate Jure Zbontar and Yann Lecun's paper of Stereo Matching by "Training a Convolutional Neural Network to Compare Image Patches".
Environement: Keras with the backend of Theano
- preprocess.py is to get the patches. This should be a brute-force way to crop the images. We create 11*11 pixel windows and train the neural networks on these training samples.
- mc-cnn-rebuild.ipynb is the first version of our neural network. We merge two CNN subnets and apply fully-connected layers on training patches of 11*11 and 9*9. testCNN.ipynb is where the disparity map is derived and the whole image is tested based on the network which has been trained. The fully-connected layer is converted into CNN layers equally to enable the different input dimensions.
- mc-cnn-rebuildONE.ipynb is a second version of our neural network. Fully-connected layers are replaced with CNN layers (same as the test network).
This is a code by Qingwei Wu and Shixin Li.