Epic Kitchens and dynibar

[f00losopher]

Hi, First of all, thanks so much for sharing your outstanding work! I would love to hear some of your Intuition and thoughts about handling egocentric data:

1. Do you think the model will fare well with the co-linear motion? This is very much the case in this dataset.
2. The dataset contains many frames and is rather long. What is a good approach for frame selection for detail and video length maximization?
3. Is a Pinehole camera model good enough for a GoPro? Should I use an open-cv one?

Sorry for the many questions, but I want to avoid as many pitfalls as possible so I can maximize your work capabilities :)

Thanks so much!

[zhengqili]

Hi,

1. I think it should work for most of moving and static regions, but might introduce artifacts around disocculusion from my observations. I am trying some tricks to fix this issue currently and update some code if this works well.
2. You can choose which frames to use for source views based on the technique from my previous paper (section 4.1 of https://arxiv.org/pdf/1904.11111.pdf), which is based on point covisibility and baselines between two views.
3. For go-pro camera, you might need to use fish-eye cameras and undistort it before training the model