D-CLOT: Double Closed Loop Optimal Transport for Unsupervised Action Segmentation
A novel technique known as D-CLOT (Double Closed Loop Optimal Transport) has been introduced to enhance unsupervised action segmentation, tackling a significant drawback of current optimal transport (OT)-based methods. This approach is built upon the newly developed CLOT framework, which improves frame embeddings by utilizing estimated segment embeddings. D-CLOT further refines action prototypes from these enhanced frame embeddings, addressing the 'representation–prototype inconsistency' that negatively affects performance, particularly during ambiguous transitions and for rare or short actions. It incorporates a graph-constrained module that maintains the local neighborhood geometry of the encoder output while regularizing the OT-refined representations. Periodic action-embedding refinement re-anchors the prototypes. This research, relevant to computer vision and video understanding, is accessible on arXiv (arXiv:2608.05877) and was announced as a cross-type submission.
Key facts
- D-CLOT is a new method for unsupervised action segmentation.
- It builds on the CLOT framework.
- It addresses representation–prototype inconsistency.
- It re-estimates action prototypes from refined frame embeddings.
- It uses a graph-constrained module to preserve local neighborhood geometry.
- It includes an action-embedding refinement step.
- The paper is available on arXiv with ID 2608.05877.
- The method targets ambiguous transitions and short or infrequent actions.
Entities
Institutions
- arXiv