Consistency and Diversity Induced Human Motion Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Data Mining

Background:

• Subspace clustering is vital for human motion segmentation but often lacks prior knowledge, leading to suboptimal results.
• Existing methods struggle to effectively integrate domain-specific and invariant properties for improved segmentation.
• Bridging the domain gap in transfer learning remains a challenge for accurate human motion analysis.

Purpose of the Study:

• To propose a novel Consistency and Diversity induced human Motion Segmentation (CDMS) algorithm.
• To enhance human motion segmentation by leveraging multi-level feature spaces and transfer subspace learning.
• To address limitations of existing methods by incorporating prior knowledge and reducing domain discrepancies.

Main Methods:

• Factorizes source and target data into distinct multi-layer feature spaces for transfer subspace learning.
• Employs multi-mutual consistency learning to minimize the domain gap between source and target data.
• Introduces a Hilbert Schmidt Independence Criterion (HSIC) constraint for diverse multi-level subspace representations and an enhanced graph regularizer for temporal correlations.
• Solves the model efficiently using the Alternating Direction Method of Multipliers (ADMM).

Main Results:

• The CDMS algorithm effectively captures multi-level information by conducting transfer subspace learning across different layers.
• Multi-mutual consistency learning successfully reduces the domain gap, enabling simultaneous exploration of domain-specific and invariant properties.
• The HSIC constraint ensures diversity in multi-level representations, boosting transfer learning performance by exploring complementarity.
• An enhanced graph regularizer preserves temporal correlations in learned representations.

Conclusions:

• The proposed CDMS algorithm significantly improves human motion segmentation compared to state-of-the-art methods.
• The integration of consistency, diversity, and temporal correlation preservation leads to more robust and accurate motion segmentation.
• The method demonstrates effectiveness on public human motion datasets, highlighting its practical applicability.