Computer vision expert needed for modify one anomaly detection architecture

i was following this code [login to view URL] in training time i tried to change I3D or TC3D by replacing C3D which used in this work but i don't have any good results. Recently i have found social mil paper (which i have uploaded) you might see they use some other brunch to generate social force map for extracting crowd interaction. my goal is similar, so instead of using social MIL learning gan/semantic segmentations/ attention network or saliency network? also i saw some recent work from slow fast network by facebook + tiny video network from google and keeping the last layer (global pool) "[login to view URL]" please be noted that i don't care abut accuracy.