Objective Classes for Micro-Facial Expression Recognition

Micro-attention for micro-expression recognition

“…Since the dataset settings in HDE and CDE are the same with what used in MEGC 2018, we directly used the baseline results achieved with LBP-TOP, HOOF and HOG3D [39], the results from study [32] as well as the state-of-the-art one [17] that won the challenge. For the detailed parameter settings, please refer to the original papers.…”

“…The same transfer learning procedure is applied for all the deep learning based methods [32] [17] [29]. In traditional LOSO on the three databases, in order to perform a fair comparison, we only compare our method with the two deep learning based approaches [17] [29] because the other methods [39][32] are either not originally tested with the three databases separately or not using the traditional LOSO.…”

“…A summary of these three databases is given in Table 1. To avoid the category bias between databases, we apply two strategies to regroup the data in each dataset respectively: i) when CASME II and SAMM are used together for cross-database validations, video clips in each database are regrouped into 5 emotion types based on the intensity distribution of AUs [38] according to the study [39]. The 5 emotions are happiness, surprise, anger, disgust and sadness, while data originally labelled as fear and others are not used; ii) when the three databases are used separately, data in CASMEII and SAMM are also regrouped into the 5 categories as stated while the category of data in SMIC stay unchanged.…”

“…With SAMM as the test set and CASME II as the training set, the residual network with proposed micro-attention achieved WAR of 0.584 and 0.341. The average WAR and UAR achieved by baseline methods [39], previous studies [17] [29] [32] and our method in two evaluation folds are summarized in Table 2. The dataset listed in each column is used as test set while another is used for training.…”

“…Still, our method yields the best performances on the three databases. [39]. In this subsection, we make efforts to interpret the modeling behavior of our method through feature visualization, which is usually employed to explain the decision of a model given its focus paid to the input data.…”