DeepMVS: Learning Multi-View Stereopsis


Po-Han Huang1   Kevin Matzen2   Johannes Kopf2   Narendra Ahuja1   Jia-Bin Huang3

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



Results on ETH3D datasets

Below we show the full qualitative results on ETH3D datasets of various MVS algorithms, including DeMoN[1], PMVS[2] (via VisualSFM), MVE, COLMAP, and our approach. All the algorithms are evaluated with the image resolution being 810×540 pixels. The zip file containing all the results is also provided.

[1]: Since DeMoN only works with images with specific focal length and resolution, the images are cropped and resized before being fed into their network. See our paper for more details.
[2]: Since PMVS generates a point cloud directly, the disparity maps shown here are generated by projecting all the points back to each view with splatting size of 3×3 pixels.

 
Method: