In simple words: CNNs handle frame content; sequence models handle change over time.
Two kinds of information
A CNN can extract features from each video frame. An RNN can read the ordered sequence of those features to model how a scene changes. This is one way to combine visual and temporal information.
Other media
Speech, text, video, and images can also be combined in other architectures. The key question is whether the model needs spatial patterns, sequence order, or both.
See the flow
frames→CNN features→RNN sequence→action label
Simple example
A CCTV clip shows a person near a bag in several frames. Frame features describe objects; sequence modelling helps describe the action.
Remember: CNNs handle frame content; sequence models handle change over time.