3. Categories of emotion

SD [32], UT DALLAS [33] and SMIC [34], and some that deal with human to human interac-

Databases produced by observing human-computer interaction on the other hand are a lot less common. The best representatives are the AIBO database [23], where children are trying to give commands to a Sony AIBO robot, and SAL [11], in which adults interact with an artificial

Even though induced databases are much better than the posed ones, they still have some problems with truthfulness. Since the emotions are often invoked in a lab setting with the supervision of authoritative figures, the subjects might subconsciously keep their expressions

Spontaneous emotion datasets are considered to be the closest to actual real-life scenarios. However, since true emotion can only be observed, when the person is not aware of being recorded [30], they are difficult to collect and label. The acquisition of data is usually in conflict with privacy or ethics, whereas the labelling has to be done manually and the true emotion has to be guessed by the analyser [25]. This arduous task is both time-consuming and erroneous [13, 38], having a sharp contrast with posed and induced datasets, where labels are either

With that being said, there still exist a few databases out there that consist of data extracted from movies [39, 40], YouTube videos [41], or even television series [42], but these databases have inherently fewer samples in them than their posed and induced counterparts. Example

tion like the ISL meeting corpus [35], AAI [36] and CSC corpus [37].

predefined or can be derived from the elicitation content.

images from these databases are in Figures 3–5 respectively.

Figure 3. Images of movie clips taken from the AFEW database [39, 40].

chat-bot.

in check [25, 30].

42 Human-Robot Interaction - Theory and Application

2.3. Spontaneous

The purpose of a database is defined by the emotions represented in it. Several databases like CK [27, 44], MMI [45], eNTERFACE [46], NVIE [47] all opt to capture the six basic emotion types: anger, disgust, fear, happiness, sadness and surprise as proposed by Ekman [48–50]. In the tables, they are denoted as primary 6. Often authors tend to add contempt to these, forming seven primary emotions and often neutral is included. However, they cover a very small subcategory of all possible emotions, so there have been attempts to combine them [51, 52].

Several databases try to just categorise the general positive and negative emotions or incorporate them along with others, e.g. the SMO [41], AAI [36], and ISL meeting corpus [35] databases. Some even try to rank deception and honesty like the CSC corpus database [37].

Apart from anger and disgust within the six primary emotions, scientists have tried to capture other negative expressions, such as boredom, disinterest, pain, embarrassment and depression. Unfortunately, these categories are harder to elicit than other types of emotions.

TUM AVIC [53] and AVDLC [12] databases are amongst those that try to label levels of interest and depression while GEMEP [31] and VAM [43] attempt to divide emotions into four quadrants and three dimensions, respectively. The main reason why most databases have a very small number of categories (mainly, neutral and smile/no-smile) is that the more emotions added, the more difficult they are to label and also more data is required to properly train a model.

Relatively newer databases have begun recording more subtle emotions hidden behind other forced or dominant emotions. Among these are the MAHNOB [51] database, which focuses on emotional laughter and different types of laughter, and others that try to record emotions hidden behind a neutral or straight face like SMIC [34], RML [54], Polikovsky's [55] databases.

One of the more recent databases, the iCV-MEFED [52, 56] database, takes on a different approach by posing varying combinations of emotions simultaneously, where one emotion takes the dominant role and the other is complimentary. Sample images can be seen in Figure 6.
