whats is the difference between the two pretrained models you supply?

[liorzamir]

I see that sky regions in the "Pretrained model for outdoor Internet Photos" tend to look quite good, how was this achieved? Reading the paper it seems sky was not part of the ground truth...

[liorzamir]

In the new sample training code it indeed seems that sky was treated with another special term encouraging far sky. Also it seems the ordinal depth term differs somewhat from the description in the paper

[stalin18]

@liorzamir : Where is the new sample training code? Which code you are referring to?

[liorzamir]

have a look at the project page: http://www.cs.cornell.edu/projects/megadepth/ in the following section they have a link: "Simple demo training code" though I am not sure it is operable code

[xuxy09]

I see that sky regions in the "Pretrained model for outdoor Internet Photos" tend to look quite good, how was this achieved? Reading the paper it seems sky was not part of the ground truth...

Have you figured out how they trained the model for outdoor photos? Did they use additional data? Or did they use a better training method? Thanks

[liorzamir]

They added the following remark the Megadepth page: "Note: for clarification, this model is used for more general purposes. We trained the network on top of DIW pretrained weights. Hence, it may have better performance than what was described in the paper." My guess is that sky regions might be inherited from its ancestor (DIW)