Comment by g_airborne

This is very cool, I’ll be studying your implementation of I3D. Did you ever attempt to train I3D end-to-end as done in the Quo Vadis paper? And it so, did you get comparable Top1/Top5 accuracy?