Fig. 7: Siamese contrastive network architecture.
Skeleton data extracted from gait videos recorded from the left and right perspectives are fed into two identical backbone networks, B1 and B2, which share the same weights. G represents the spatial topology of the data, capturing the spatial dependencies among nodes. These dependencies are propagated through the network via graph convolution operations. Each backbone network extracts feature vectors, denoted as f1 and f2, which are subsequently concatenated into a unified vector, f. This composite vector is then processed through fully connected layers followed by a softmax layer, yielding a probabilistic distribution across three distinct classes. The right part shows the structure of the spatial-temporal graph convolutional (ST-GCN) network.
