One-Shot Action Recognition via Multi-Scale Spatial-Temporal Skeleton Matching.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Collecting and annotating large-scale skeleton action data is challenging, limiting traditional action recognition models.
• Existing one-shot skeleton action recognition methods often overlook the spatial and temporal dynamics inherent in skeleton data by relying on direct feature vector comparison.

Purpose of the Study:

• To develop a novel technique for one-shot skeleton action recognition that effectively utilizes multi-scale spatial-temporal features.
• To address the limitations of existing methods by incorporating spatial structures and temporal orders for more robust action recognition.

Main Methods:

• Representing skeleton data across multiple spatial and temporal scales.
• Implementing multi-scale matching to capture scale-wise semantic relevance simultaneously.
• Employing cross-scale matching to handle variations in motion magnitude and speed by analyzing inter-scale relevance.

Main Results:

• The proposed method demonstrates superior performance in one-shot skeleton action recognition.
• Consistent and significant performance improvements were observed over state-of-the-art (SOTA) methods.
• Experiments were validated on three large-scale datasets: NTU RGB+D, NTU RGB+D 120, and PKU-MMD.

Conclusions:

• The multi-scale spatial-temporal feature matching approach is highly effective for one-shot skeleton action recognition.
• This technique offers a robust solution for scenarios with limited training data.
• The method significantly advances the field by providing a more comprehensive analysis of skeleton data.