DeepMVS: Learning Multi-View Stereopsis

Po-Han Huang1Kevin Matzen2Johannes Kopf2Narendra Ahuja1Jia-Bin Huang3

1University of Illinois, Urbana Champaign   2Facebook   3Virginia Tech

Comparisons with Conventional MVS Algorithms

Below are the comparisons between our approach and other conventional multi-view stereo algorithms evaluated on ETH3D Dataset.

Reference Ground Truth PMVS MVE COLMAP (filtered) COLMAP (unfiltered) Our Result

Comparisons with DeMoN on Image Pairs

Below are our results on DeMoN's testing dataset, which consists of image pairs. Our algorithm is designed for sequences with more images, but is still capable of generating disparity maps for image pairs.

Image1 Image2 Ground Truth DeMoN Our Result