Repository files navigation

A Lasagne and Theano implementation of the paperConvolutional neural network architecture for geometric matching by Ignacio Rocco, Relja Arandjelović, and Josef Sivic.

Download thePascal VOC dataset.

(Optional) Download theProposal Flow dataset.

Download the pre-trained weights forVGG16.

(Optional) Download thelearnedweightsif you don't want to train your own model from scratch (NOTE: These weightswon't reproduce the figures from the paper, because I haven't implemented thethin-plate-spline transform yet).

Get the code necessary for generating random transformation matrices fromthis repo.

This is a work-in-progress. Pull requests are welcome. Contact me if you run into issues using the code.

The thin-plate-spline has not yet been implemented. The model has not beentrained properly yet, either. The images below were taken from the validationset after training for 300 epochs (about 17 hours on a TITAN X). The image onthe left is the center crop, the image in the middle is the result of applyingthe ground-truth transformation to the center crop, and the image on the rightis the result of applying the predicted transformation to the warped image (inother words, the pose of the rightmost image should resemble that of theleftmost image).

Similarly, the images below are from the Proposal Flow dataset: