Siamese Network with Deconvolution for Matching Cost Computation
Hang Li, Yan Song, Yongxiong Wang, Ming Li · 2018
An efficient Siamese network (SiaNet) with deconvolution is proposed for matching cost computation of the stereo estimation. The architecture of such a proposed SiaNet with deconvolution which is established by learning a similarity measure on small image patches, is a multi-class classification with the classes corresponding to all possible disparities. By adding some deconvolution layers, the architecture of our SiaNet can obtain a larger receptive field as a greater descriptor for the output of network branches. Then, a new loss function based on cross-entropy is developed by taking all the pixels in the receptive field with the corresponding ground-truth patch into account. As such, the performance can be effectively improved by considering the more information of the input patches. Finally, some experiments for matching network on the KITTI 2012 and KITTI 2015 are utilized to demonstrate the validity of the proposed SiaNet with deconvolution.