Self-Supervised Skeleton Representation for Imbalanced Ethnic Dance Motion Analysis

Abstract

Digitally preserving traditional ethnic choreography faces many challenges, including severe intraclass kinematic variance, inconsistent sensor frame rates, and class imbalance. This paper seeks to advance the state-of-the-art in human movement representation by structuring cohesive, semantically aligned feature spaces with unannotated spatiotemporal data. We proposed MADALA-SSL, a selfsupervised multi-task learning framework that jointly optimizes a label-aware supervised contrastive loss, a temporal order prediction task, and a masked motion reconstruction task. We tested our framework on an imbalanced Bugis-Makassar Paduppa dance pose skeleton dataset. This dataset includes 144 sample clips of 6 unique motions captured by a heterogeneous setup of lateral 4 K 60 FPS and frontal 1080p 25 FPS, with a consistent temporal frame rate of 60 FPS. Using a frozen linear probing with a 5-fold cross-validation, MADALA-SSL reached a maximum accuracy of 89.29% and a macro F1 of 0.7458. Topological assessments of the data shape with a mean silhouette score of 0.4790 and a DBS of 1.2242, revealed an overall cohesive topology of the data despite some kinetic leakage. This was due to low-resource Massimang sequences merging into the Ammula Gau class, which has a highly imbalanced class distribution, caused by indistinguishable starting stances of the movements. Integrating self-paced learning loops of cross-modal audio-kinetic rhythmic synchronization is a promising future direction for this work.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint