Can Expression Sensitivity Improve Macro- and Micro-Expression Spotting in Long Videos?

Spotting people's expressions is pivotal, as it directly reflects emotions, particularly those underlying feelings and intentions that may not be expressed verbally. Therefore, detecting macro- and micro-expressions play a critical role in psychological analysis by providing an understanding of nuanced emotional states. This detection process specifically involves spotting intervals within a long video that contain either macro- or micro-expressions, distinguishing these crucial moments from other time-scale facial movements. Since successful outcomes in the task of micro-expression recognition have yield, this paper proposes an algorithm that leverages these outcomes to enhance the performance of spotting both macro- and micro-expressions using an expression-sensitive model. This study introduces a novel and effective approach that underscores the importance of integrating recognition and detection tasks of facial expressions. Moreover, repurposing datasets initially designed for micro-expression recognition tasks for detection tasks has broadened their scope, thereby enhancing their overall utility. The proposed algorithm employs a pre-trained CNN model and uses a unique confidence value-based mechanism to label the training samples. This approach not only achieved superior performance compared to other advanced methods but also demonstrated an exceptional balance across various databases, between macro- and micro-expressions, and in terms of precision and recall.