Asymmetric Learning for Stereo Matching Cost Computation

Zhongjian MA  Dongzhen HUANG  Baoqing LI  Xiaobing YUAN  

IEICE TRANSACTIONS on Information and Systems   Vol.E103-D   No.10   pp.2162-2167
Publication Date: 2020/10/01
Publicized: 2020/07/13
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2020EDP7002
Type of Manuscript: PAPER
Category: Artificial Intelligence, Data Mining
stereo matching,  asymmetric convolutions,  feature extraction,  CNN,  

Full Text: PDF>>
Buy this Article

Current stereo matching methods benefit a lot from the precise stereo estimation with Convolutional Neural Networks (CNNs). Nevertheless, patch-based siamese networks rely on the implicit assumption of constant depth within a window, which does not hold for slanted surfaces. Existing methods for handling slanted patches focus on post-processing. In contrast, we propose a novel module for matching cost networks to overcome this bias. Slanted objects appear horizontally stretched between stereo pairs, suggesting that the feature extraction in the horizontal direction should be different from that in the vertical direction. To tackle this distortion, we utilize asymmetric convolutions in our proposed module. Experimental results show that the proposed module in matching cost networks can achieve higher accuracy with fewer parameters compared to conventional methods.