Developing Feature Extraction and Clustering Approaches for Temporal-Relation-Based Action Prediction

In this work, we propose a weakly supervised framework for action segmentation that automatically discovers temporal structures without predefined cluster numbers. The method combines DP-Means to estimate temporal anchors, Pseudo-Label Ensembling (PLE) to generate reliable pseudo-labels from complementary clustering strategies, Iterative Clustering (IC) to refine boundaries and propagate labels to uncertain regions. This design enhances temporal consistency, boundary precision, robustness under sparse supervision. Experiments on GTEA and 50Salads show that the proposed approach consistently outperforms individual clustering methods and conventional pseudo-labeling techniques while substantially reducing annotation costs.