Wide context learning network for stereo matching

作者：

Highlights：

• Performance of an end-to-end convolutional neural network without conventional stereo methods.

• Find features on local and wide context using convolutional networks without losing spatial information.

• Reducing computational cost by evaluating feature maps using cosine similarity function before regressing disparity.

• Producing accurate disparity on ill-posed regions such as slanted plane, textureless, flat, repetitive regions, discontinuities, and occlusions.

摘要

•Performance of an end-to-end convolutional neural network without conventional stereo methods.•Find features on local and wide context using convolutional networks without losing spatial information.•Reducing computational cost by evaluating feature maps using cosine similarity function before regressing disparity.•Producing accurate disparity on ill-posed regions such as slanted plane, textureless, flat, repetitive regions, discontinuities, and occlusions.

论文关键词：Stereo matching,3D reconstruction,Matching cost,Cost aggregation,Convolutional neural network

论文评审过程：Received 7 January 2019, Revised 6 July 2019, Accepted 15 July 2019, Available online 24 July 2019, Version of Record 29 July 2019.

论文官网地址：https://doi.org/10.1016/j.image.2019.07.008