ADR-MVSNet: A cascade network for 3D point cloud reconstruction with pixel occlusion

作者：

Highlights：

• The adaptive depth reduction module learns different depth ranges for each pixel.

• The occlusion map reflects the visibility of pixels in different views.

• The outliers in depth map are filtered by multi-scale photometric consistency.

摘要

•The adaptive depth reduction module learns different depth ranges for each pixel.•The occlusion map reflects the visibility of pixels in different views.•The outliers in depth map are filtered by multi-scale photometric consistency.

论文关键词：3D point cloud reconstruction,Multiview stereo,Deep neural network,Cost volume

论文评审过程：Received 4 June 2021, Revised 24 November 2021, Accepted 28 December 2021, Available online 2 January 2022, Version of Record 10 January 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108516