Convolutional neural network and adaptive guided image filter based stereo matching

This paper presents a novel stereo matching algorithm based on convolutional neural network (CNN) and adaptive guided image filter. Firstly, we trained a convolutional neural network through learning a similarity measure on small image patches to initialize the matching cost. This method can extract the characteristics of the pictures automatically and precisely, and has strong robust against radiometric variations. Then, we aggregate the cost volume with guided image filter whose support window is adaptive rectangular instead of the traditional fixed support window. The variation of the window's kernel is generated by the local spatial distance, color similarity and gradient so that less occluded points will be included in the support region. Moreover, we adopt integral image and box filter to further speed up the computation of this step. At last, we evaluate our method on the Middlebury and show that it preserves the edges well and outperforms the traditional methods greatly.