Wrong matching in demo_match

[npmhung]

Thank you for your great work.

To get the uploaded result, I modified the code a bit so that all certainty values>0.05 are changed to 1 as follow:
certainty[certainty>0.05] = 1.

The 2 corresponding frames are also shown in the top row for comparison.

If you look at the red car near the bridge in the first image, you can see that the warped version somehow moves it toward the bridge.

This effect seems to happen to dynamic/occluded regions only. The warped image in the following example is pretty good.

Could you think of any potential reasons causing this problem?

Thank you
@Parskatt

[npmhung]

One more example with the matching.

[Parskatt]

Hi!

The matcher is trained only on static objects. So it's likely that dynamic objects get a low certainty and don't match well. It's a limitation of the current training. Could be fixed by having dynamic data for training.