Self-Supervised Sub-Action Parsing Network for Semi-Supervised Action Quality Assessment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised Action Quality Assessment (AQA) faces challenges in leveraging limited labeled data alongside abundant unlabeled data.
• Developing consistent representations of action sequences is crucial for bridging the gap between labeled and unlabeled samples in semi-supervised AQA.

Purpose of the Study:

• To propose a novel Self-supervised sub-Action Parsing Network (SAP-Net) for effective semi-supervised AQA.
• To enhance the learning of consistent semantic representations for action sequences.

Main Methods:

• Employed a teacher-student network structure for learning consistent representations.
• Utilized actor-centric region detection and pseudo-label generation in the teacher branch.
• Implemented a self-supervised sub-action parsing solution for fine-grained sequence analysis.
• Applied group contrastive learning with pseudo-labels to capture motion-oriented features.

Main Results:

• SAP-Net demonstrated superior performance compared to existing state-of-the-art semi-supervised methods.
• The approach successfully learned discriminative action features by bridging labeled and unlabeled data.
• Consistent motion-oriented action features were effectively captured.

Conclusions:

• SAP-Net offers a robust solution for semi-supervised AQA by effectively exploiting unlabeled data.
• The proposed method significantly advances the field of action quality assessment.