Prototype Contrastive Learning

Prototype contrastive learning leverages the power of contrastive learning to learn robust representations from skeletal data, focusing on identifying and contrasting "prototypes"—highly representative features—within different skeletal configurations. Current research emphasizes hierarchical and multi-level graph representations of skeletons, often incorporating transformers and incorporating various augmentation strategies (including attack-based methods) to generate informative positive and negative sample pairs for contrastive learning. This approach shows promise for improving performance in tasks like person re-identification and action recognition, particularly in scenarios with limited labeled data, by learning more generalizable and discriminative features from skeletal information.