Bad Results Witnessed on RPG Dataset Since 88bb052

[Ethan-Zhou]

Recently, some people reported that they could not obtain ideal results on RPG dataset. As far as I know, this is very likely caused by the modification in commit 88bb052. Please refer to the commit message for further explanation.

For people who are working on small scale environment (with a small baseline configuration) or want to re-generate ideal results on RPG dataset, a quick-hand solution is to checkout an older commit earlier than 88bb052, e.g., 538b576.

I will try to find a unified termination criteria for the inverse depth estimation problem in both small-scale and large-scale data.