Classfication and Box Regression based on RoI Pooling and RoI tracking ?

[sharma-n]

Hi @Feynman27 , Thanks for the effort of translating Detect to Track to pytorch!

I was wondering whether to predict the classification and box regression of the frame t, you use only the RoI poolings from frame t? I think in the paper it appears that they combine all the RoI pooling output from both time t and t+tau, and also the RoI Tracking output. However, in the RFCN code (Line 108) it seems that you calculate the classification and regression from just the output of one leg of the network.

Have I understood something wrong? Thanks!