MC-HDCNN: Computing the stereo matching cost with a hybrid dilated convolutional neural network

Abstract

Designing a model to quickly obtain an accurate matching cost is a vital problem in the stereo matching method. We present an algorithm called MC-HDCNN, which is based on hybrid dilated convolution neural network, for computing matching cost of two image patches. HDCNN uses the dilated convolution of the series to obtain a larger receptive field, while avoiding the “gridding” effect and ensuring the integrity of the receptive field. In addition, by adding batch normalization layer after each layer of the convolution, the gradient dispersion in the backward propagation and the generalization of the network can be improved effectively. We evaluate our method on the KITTI stereo data set. The results show that the proposed algorithm has certain advantages in accuracy and speed.