Temporal sampling strategy and computational cost

[zhenhuat]

Hello, thank you very much for sharing. I have two issues:
（1）Where is the temporal sampling strategy in your code? It seems to be the most important trick of your model.
（2）I can't train the 9-frame model on the P40 server （with 24G memory). The computational cost seems much greater than the values (FLOPs) reported in your paper.