The deep features are propagated using a bilinear warping function:
Does this video belong to a specific like ImageNet VID, or are you looking to implement this on a custom real-time stream ? 0guogcfcb4q156ug2eqlg_source.mp4
): The model runs a full forward pass through the feature network ( Nfeatcap N sub f e a t end-sub ) to get feature maps A lightweight FlowNet ( Nflowcap N sub f l o w end-sub ) calculates the displacement field ( Mi→kcap M sub i right arrow k end-sub ) between the current frame and the last keyframe. The deep features are propagated using a bilinear