Hi! Thank you for releasing this wonderful work. I have a question on how RFR operates:
Since the dependencies of F_t^i are the items to its LEFT and TOP (like this drawing), and we get shallow features for all clips (t=0,1,...T/N -1) at layer i=0, how do we calculate features for clip t=0 at all layers (i=1,2,...L-1) in order to start the recurrency for later clips?
For example, to calculate F(i=1, t=1), it depends on F(i=0, t=0), F(i=0, t=1) (these two I understand are the shallow features to begin with), and F(i=1, t=0) (which I'm not sure where it comes from).
Thank you in advance! I'd appreciate your help in explaining!
Hi! Thank you for releasing this wonderful work. I have a question on how RFR operates:
Since the dependencies of F_t^i are the items to its LEFT and TOP (like this drawing), and we get shallow features for all clips (t=0,1,...T/N -1) at layer i=0, how do we calculate features for clip t=0 at all layers (i=1,2,...L-1) in order to start the recurrency for later clips? For example, to calculate F(i=1, t=1), it depends on F(i=0, t=0), F(i=0, t=1) (these two I understand are the shallow features to begin with), and F(i=1, t=0) (which I'm not sure where it comes from).
Thank you in advance! I'd appreciate your help in explaining!