Related to #26. How much is flow actually helping us, if at all? The Modeep paper suggested that no motion features are particularly helpful (not even optical flow!), so it could be that it's not really improving our network performance.
In line with the MoDeep paper, it would also be interesting to try including only the magnitude of optical flow. It sounds like networks can overfit quite a bit on the direction of flow, which could explain why my networks have been overfitting so badly on the datasets I've tried.
Related to #26. How much is flow actually helping us, if at all? The Modeep paper suggested that no motion features are particularly helpful (not even optical flow!), so it could be that it's not really improving our network performance.
In line with the MoDeep paper, it would also be interesting to try including only the magnitude of optical flow. It sounds like networks can overfit quite a bit on the direction of flow, which could explain why my networks have been overfitting so badly on the datasets I've tried.