Abstract: ‍ ‍Emotion recognition is an essential process of natural human-computer interaction. However， the high cost of data collection and labeling has become a bottleneck in the development of emotion recognition research. It is worth exploring how to improve the recognition performance by the cross-domain or cross-task knowledge transfer in scenarios with no or limited annotations. This paper organizes and analyzes transfer learning problems in emotion recognition. Firstly， transfer learning problems are divided into two parts， which are problems for domain discrepancies and for task differences. Each part of them is further subdivided into several different situations. Then， existing works of emotion recognition in different situations are summarized respectively. In the case of scarce training resources in the target domain， other annotated datasets can be used as the source domain to train the model. During this process， the feature distributions from different domains should be aligned， or the features should be mapped to a shared space. Considering that the supervision information provided by emotion annotations is often limited， in order to further improve the recognition performance of the model， other related tasks can be introduced for joint training， or the prior semantic knowledge provided by the pre-training model and external knowledge base can be transferred to the emotion recognition task. Finally， transfer learning problems in the emotion recognition task which need more attention and exploration in the future are discussed， aiming to bring new inspiration to researchers.