基于自适应卷积的视差细化网络  

Parallax Refinement Network Based on Adaptive Convolution

作　　者：胡文辉 周泽豪 戚桢[1] 

机构地区：[1]天津工业大学,天津

出　　处：《光电子》2022年第4期147-158,共12页Optoelectronics

摘　　要：为提高视差细化的精度,本文提出一种基于自适应卷积的视差细化与采样方法。利用细化假设为视差细化引入其它可用信息,在不同阶段附加不同的信息来增强细化假设。基于自适应传播方法构建局部代价卷,并将聚合操作从空间域转换至视差域,以缓解使用大卷积窗口带来的边界模糊问题,增强在无纹理或弱纹理区域的聚合效果。同时,使用自适应卷积从相似视差平面上更新视差,进而提高视差的精度。对于上采样过程,利用视差自适应采样克服双线性插值导致的精度下降问题。在SceneFlow和KITTI2015数据集上,对算法进行验证,实验结果表明,相比原始方法,本文算法在精度方面有了明显提升,特别是在KITTI2015数据集上,端点误差(EPE)和3像素错误率指标分别提升9.7%和12.5%。To improve the accuracy of disparity refinement, this paper proposes a disparity refinement and sampling method based on adaptive convolution. The refine hypothesis is used to introduce other available information for disparity refinement, and different information is attached at different stages to augment the refine hypothesis. Based on the adaptive propagation method, the local cost volume is constructed and the aggregation operation is converted from the spatial domain to the disparity domain to alleviate the boundary blur problem caused by using large convolution windows and augmented the aggregation effect in texture-free or weakly textured areas. At the same time, adaptive convolution is used to update the disparity from similar disparity planes, which in turn improves the accuracy of disparity. For the upsampling process, disparity adaptive sampling is used to overcome the degradation of accuracy caused by bilinear interpolation. The algorithm is validated on the SceneFlow and KITTI2015 datasets, and the experimental results show that the algorithm in this paper has significantly improved in terms of accuracy compared with the original method, especially on the KITTI2015 dataset, the Endpoint Error (EPE) and 3-pixel error rate indicators have increased by 9.7% and 12.5%, respectively.

关 键 词：机器视觉 立体匹配 双目视觉 自适应卷积 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]