Bigger error than reported in paper on middlebury dataset

[za-cheng]

Hi,

When I ran the demo script on Middlebury I get an average IE of 4.5146 with pre-trained model. That's much bigger than errors reported in Table.1 of paper which are around 2.1 . Could you offer some insight in this? Thanks!