Learning Clip Representations for Skeleton-Based 3D Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• 3D action recognition is crucial for human-computer interaction.
• Existing methods struggle with complex spatial-temporal dynamics in skeleton data.
• Hand-crafted features and recurrent neural networks have limitations.

Purpose of the Study:

• To propose a new skeleton sequence representation for improved 3D action recognition.
• To develop a novel deep learning model for learning from the proposed representation.
• To demonstrate the effectiveness of the proposed approach on benchmark datasets.

Main Methods:

• Transforming 3D skeleton coordinate channels into 'clips'.
• Each clip frame encodes temporal information and spatial joint relationships.
• Utilizing a multitask convolutional neural network (MTCNN) for parallel processing of clip frames.
• Leveraging the MTCNN to explore both spatial and temporal features.

Main Results:

• The proposed clip representation effectively captures spatial structural information.
• The MTCNN successfully learns from the generated clips, extracting rich features.
• Consistent superiority demonstrated across six benchmark datasets.
• Outperformed existing techniques in 3D action recognition tasks.

Conclusions:

• The novel clip representation significantly enhances 3D action recognition.
• The MTCNN provides an effective framework for learning spatial-temporal features.
• The proposed method offers a superior alternative for skeleton-based action recognition.