Difficulty Getting this to work over Egocentric Data

[edess442]

Hello,

Did the parameters as chosen in this code here work over datasets that are egocentric, like Ego4D and EPIC Kitchens? I'm having a very difficult time getting foreground masks that don't include huge swathes of background. I'd say it's only about 20% of the time the masks include contacted objects.

[edess442]

I used a more recent flow model, changed the source of the fundamental matrix (from RANSAC to IMU-based SLAM) and decreased the timespan of the algorithm from 0.5 seconds to 0.1 seconds. This produces much better results. Not perfect, but better.