“…Specifically, handcrafted features (e.g., LBP [81,83], Gabor [22,115], HOG [36,87,104]) or deep-based features (e.g., CNN [74,115], VGG [91], ResNet [39]) are employed to enhance the node representation similar to many non-graph FER methods [20,31]. For reasoning approaches, early studies prefer to capture the relations of an individual node from predefined graph structures using tracking strategies (e.g., displacement projection [76] and DNG [90]) or general machine learning models (e.g., RF [40], RNN [22], CNN [36]). In the latest work, GCNs become one of the mainstream choices in the latest work and show state-of-the-art performances on posed and inthe-wild databases [39,74,81,101,105,113,115].…”