TVPR: A Multilayer Convolutional Neural Networks for Facial Micro- and Macro-Expression Spotting

Facial expression analysis is a vital research field that encompasses expression spotting. This process entails spotting both micro-expressions (MEs) and macro-expressions (MaEs) in videos. In this paper, we introduce the TVPR network, which utilizes TVConv and PRConv layers to capture spatial and temporal features effectively. Additionally, the use of Gaussian and linear pseudo-labels enhances expression re-labeling, improving the accuracy of apex frame localization. The proposed method's superior performance in expression spotting is demonstrated through experiments conducted on the CAS(ME)2 and SAMM Long Videos (SAMMLV) datasets. The TVPR network, along with the novel re-labeling approach, makes a substantial contribution to advancing facial expression analysis. We have exceeded the baseline methods, obtaining an overall F1-score of 0.2125. Specifically, we achieved F1-score of 0.2624 on the CAS(ME)2 dataset and 0.1884 on the SAMMLV dataset.