Beyond Pattern Variance: Unsupervised 3-D Action Representation Learning With Point Cloud Sequence.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised 3-D action representation learning from point cloud sequences is a challenging area.
• Existing methods often rely on 3-D skeleton information, limiting their applicability.
• Current contrastive learning (CL) methods struggle with high pattern variance in augmented 3-D dynamic voxel (3DV) samples.

Purpose of the Study:

• To develop an unsupervised method for 3-D action representation learning using point cloud sequences.
• To address limitations of existing CL methods in handling augmented 3DV samples.
• To improve the discriminative power of learned 3-D action features.

Main Methods:

• Utilized the 3-D dynamic voxel (3DV) descriptor to represent 3-D motion information from point cloud sequences.
• Applied contrastive learning (CL) with spatiotemporal data augmentations on 3DV samples.
• Proposed a novel feature augmentation adapted CL (FACL) approach with global and local feature learning branches.
• Fused global and local features for joint 3-D action characterization.

Main Results:

• The proposed FACL method effectively addresses high pattern variance in augmented 3DV samples.
• Experiments demonstrate superior performance of the unsupervised learning method for 3-D action feature learning.
• Outperformed state-of-the-art skeleton-based methods by 6.4% and 3.6% on NTU RGB+D 120 dataset.

Conclusions:

• The FACL approach significantly enhances unsupervised 3-D action representation learning.
• This method offers a promising alternative to skeleton-based approaches for 3-D action recognition.
• The study provides a valuable contribution to the field of 3-D action understanding.