It would be interesting to train a new model with a frame skip of zero (so each frame pair passed to the CNN consists of two identical frames). That would identify precisely how much performance is improved through use of motion information. It might also be a good way of demonstrating the contribution of this work (or even the lack thereof, if it comes to that!).
It would be interesting to train a new model with a frame skip of zero (so each frame pair passed to the CNN consists of two identical frames). That would identify precisely how much performance is improved through use of motion information. It might also be a good way of demonstrating the contribution of this work (or even the lack thereof, if it comes to that!).