Institute of psychology, Chinese Academy of Sciences

Visual and computational cognitive Laboratory

No 16, Lincui Road, Chaoyang District, Beijing




Micro-expression (ME) is a brief and subtle facial movement, which appears when an individual tries to conceal his/her genuine emotion. Whether it is to study the characteristics of human micro-expression recognition, or to develop ME spotting and recognition algorithms, it depends to a large extent on the ME database. In order to overcome the shortcomings of the previous databases (such as low ecological validity, lax analysis and coding), Fu Xiaolan's team from the Institute of Psychology, Chinese Academy of Sciences established three natural spontaneous ME databases: CASME, CASME Ⅱ and CAS(ME)².

Participants were asked to watch videos of different emotions while keeping their faces neutral during data collection. During the collection process, the subjects were asked to remain expressionless and motionless. By eliminating the interference of non-emotional factors such as light and head movement, purer ME samples were obtained. Coding the action unit (AU) of the ME samples helps to annotate the expressions objectively and accurately. Regarding the ME emotion labels, based on psychological research, the characteristics of AU, video materials, and subjective reports of subjects are considered.

Although we have three ME databases, there is a problem of small sample size. The sample size of currently public ME databases is relatively small, mainly because it is difficult to induce micro-expression and the annotation process is very time-consuming and laborious. The lack of samples limits the training of computers to realize automatic ME spotting and recognition.

There is also the problem of inconsistent emotion labels. Different databases have different emotion classification standards, which leads to the problem of label inconsistency when using various databases to train ME recognition algorithms. Therefore, some researchers often re-categorize emotions when using ME samples, such as changing the original six classes into three or four classes, and dividing all samples into positive, negative, surprised, and others. Positive expression includes happy expressions, which are relatively easy to induce and have obvious characteristics. Negative expressions include disgust, sadness, fear, anger, etc. These micro expressions are relatively difficult to distinguish, but they are significantly different from positive micro expressions. Meanwhile, surprise has no direct relationship with positive or negative expressions, and expresses unexpected emotions, which can be interpreted according to the context. Such classification can be better compatible with different databases and has better psychological support.