I was reading this paper and my question is related to it. What I am stuck at is the intuition behind using two CNNs for feature extraction. Can just RNN not be used for feature extraction as well as for capturing temporal relationship?
In theory, any neural network can be used for feature extraction since neural nets perform it by themselves.
Whether or not using CNNs vs RNNs is a different question. CNNs are typically more computationally efficient whilst RNNs are more flexible but typically slightly harder to train.