You can use either blendedmvs-trained dataset or use dtu-trained model. They are provided in release
During evaluation, since some scenes have more than 800 views, you might need limit the number of refernce views by setting --max_ref_views to a proper number (400 by default) as the program requires huge RAM which grows with the image size and the number of reference views. I tested with 32GB RAM, and it is almost all consumed using default settings, if you have more RAM, feel free to set this number (and the image size) larger.
No point cloud evaluation currently.
Please see BlendedMVS_scenes for a quick lookup of various scenes.
Point clouds are provided in release.
Some examples:
I'm very impressed by this scene. The unevenness of the sculpture is reconstructed to a degree that it looks almost like if it were carved by a veteran artist's hand.
A youtube video showing the point cloud in interactive view.