Classification of egocentric and exocentirc videos using a 3DCNN, with pretrained weights from the 1M Sports dataset; Also includes an option to use optical flow.
I use a method to have a shared 3D-CNN backbone to classify videos as ego and exocentric videos and this backbone also supports the task of performing activity recognition on the video.