D-CLOT Refines Unsupervised Action Segmentation via Double Closed Loops
August 7, 2026
D-CLOT addresses representation-prototype inconsistency in unsupervised action segmentation by re-estimating action prototypes from refined frame geometry. It uses a graph-constrained module to regularize frame and segment representations while preserving local neighborhood geometry.
HOW THIS AFFECTS YOU
●
researcherYou can improve segmentation accuracy around ambiguous transitions by ensuring latent prototypes are consistent with refined embeddings.