Explore Efficient Local Features from RGB-D Data for One-Shot Learning Gesture Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• RGB-D sensors enable advanced gesture recognition research.
• One-shot learning offers an efficient approach requiring minimal training data.

Purpose of the Study:

• To review one-shot learning for RGB-D gesture recognition.
• To propose a novel spatiotemporal feature, Mixed Features around Sparse Keypoints (MFSK).
• To address challenges in one-shot gesture recognition and suggest future research.

Main Methods:

• A thorough review of one-shot learning for RGB-D gesture recognition.
• Development and application of the novel MFSK feature.
• Data augmentation by synthesizing temporal scales to address limited training samples.
• Evaluation on the Chalearn Gesture Dataset (CGD) and other RGB-D datasets.

Main Results:

• The proposed MFSK feature demonstrates robustness and invariance to scale, rotation, and partial occlusions.
• The method significantly outperforms existing approaches on challenging subsets of the CGD.
• Promising results achieved on non-one-shot RGB-D datasets using cross-validation and one-shot learning.

Conclusions:

• MFSK is a highly effective feature for one-shot gesture recognition from RGB-D data.
• The proposed augmentation strategy improves performance with limited training samples.
• The approach shows strong generalization capabilities across different datasets and learning scenarios.