Flipped flow maps in the flow loss

[sanjayharesh]

Hi Gengshan,

Thanks for the great work.

I noticed that you flip the flows before saving in autogen.py https://github.com/google/lasr/blob/29d8759354f853119276f41504d6b527fd5484e5/preprocess/auto_gen.py#L173-L178

Is there a good reason to do this? An unintended consequence is that it leads to flipped flows being loaded at training time and the flow loss ends up being wrong.

Here's an example of flow error being logged while running lasr on Camel example. As you can see, the ground truth flow here is flipped

[gengshan-y]

Hi, it doesn’t save flipped flow image since write_pfm() internally flips it here. Can you check whether the color-coded flow images are flipped?

[sanjayharesh]

Yes, there's a corresponding flip operation in readPFM as well, https://github.com/google/lasr/blob/29d8759354f853119276f41504d6b527fd5484e5/third_party/ext_utils/util_flow.py#L79

So maybe the initial flipping is not needed?

[gengshan-y]

Yeah, I believe the flip operation is not necessary within the the pfm functions.