End-to-End Disparity Estimation with Multi-granularity Fully Convolutional Network

Abstract

Disparity estimation is a challenging task in the field of computer stereo vision. In this paper, we propose a multi-granularity fully convolutional network architecture for end-to-end dense disparity estimation. First, we use single well-pretrained residual network for extraction of multi-granularity and multi-layer features. Second, correlation layers at three different granularities are used to gain hierarchical matching cues between left and right feature maps. Third, we conduct concatenation-deconvolution operations to output disparity maps. Finally, the experimental results show that our method achieves state of the art results, taking the second place on the KITTI Stereo 2012 task.