Learning a Target Sample Re-Generator for Cross-Database Micro-Expression Recognition

Abstract

In this paper, we investigate the cross-database micro-expression recognition problem, where the training and testing samples are from two di erent micro-expression databases. Under this setting, the training and testing samples would have di erent feature distributions and hence the performance of most existing microexpression recognition methods may decrease greatly. To solve this problem, we propose a simple yet e ective method called Target Sample Re-Generator (TSRG) in this paper. By using TSRG, we are able to re-generate the samples from target micro-expression database and the re-generated target samples would share same or similar feature distributions with the original source samples. For this reason, we can then use the classi er learned based on the labeled source samples to accurately predict the micro-expression categories of the unlabeled target samples. To evaluate the performance of the proposed TSRG method, extensive cross-database micro-expression recognition experiments designed based on SMIC and CASME II databases are conducted. Compared with recent state-of-the-art cross-database emotion recognition methods, the proposed TSRG achieves more promising results. * Yuan Zong is also with the

Toward Bridging Microexpressions From Different Domains

“…Recently, the problem of cross-domain microexpression recognition has gradually drawn the researchers' attentions. In the work of [17], Zong et al first investigated unsupervised cross-database microexpression recognition by proposing a target sample regenerator (TSRG) method. This method targets at learning a sample regenerator to regenerate the source and target microexpression samples such that the regenerated source and target samples would have the same or similar feature distributions.…”

“…Following the work of [17], the face images of the video clips from CASME II are cropped and transformed to 308 × 257 pixels, while for SMIC databases, we crop and then transform the images of microexpression samples into 170 × 139 pixels. As for performance evaluation metrics, we report the experimental results in terms of both weighted average recall (WAR) and unweighted average recall (UAR), which are widely used in the cross-domain speech emotion recognition research [19].…”

“…As for performance evaluation metrics, we report the experimental results in terms of both weighted average recall (WAR) and unweighted average recall (UAR), which are widely used in the cross-domain speech emotion recognition research [19]. WAR is actually the normal recognition accuracy, while UAR is defined as the mean accuracy of each For comparison purposes, we choose nine representative DA methods, that is, KMM [23], [24]; KLIEP [23], [25]; uLSIF [23], [26]; transfer component analysis (TCA) [50]; GFK [31]; subspace alignment (SA) [51]; STM [27], [28]; transfer kernel machine (TKL) [52]; and TSRG [17] to conduct the experiments under the same protocol as our TTRM + ASSM method. Meanwhile, we use the SVM without any DA to conduct the experiments to serve as the baseline.…”

“…For this reason, the cross-validation method is not practical to determine the tradeoff parameters in the experiments. Consequently, to offer a fair comparison among all methods, in this paper, we use the widely used parameter space search strategy [17], [52]- [55] in unsupervised DA experiments for these methods and report the best results which correspond to the optimal parameters with a preset parameter space. The details of the parameter setting for these DA methods are as follows.…”

“…As a typical domain adaptation (DA) problem, the crossdomain microexpression recognition can be roughly divided into two cases according to the number of the provided labeled samples from the target domain, that is, the semisupervised case and unsupervised case [17]. Table I illustrates the detailed label information provided in these two cases.…”