Comparing Handcrafted Features and Deep Neural Representations for Domain Generalization in Human Activity...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) models lack generalization across diverse domains (subjects, devices, datasets).
• This limits real-world applicability of current HAR approaches.
• Deep neural networks are increasingly used, necessitating a comparison with traditional methods.

Purpose of the Study:

• To compare handcrafted and deep learning representations for Human Activity Recognition.
• To evaluate model performance in Out-of-Distribution (OOD) settings.
• To assess generalization capabilities across multiple domains.

Main Methods:

• Comparison of handcrafted and deep learning features using homogenized public datasets.
• Validation of three distinct OOD settings using various metrics.
• Experimental verification of model performance under increasing distribution shifts.

Main Results:

• Deep learning models initially show superior performance.
• Handcrafted features demonstrate better generalization as the distribution shift increases.
• Performance reversal observed between deep learning and handcrafted features in OOD scenarios.

Conclusions:

• Handcrafted features show potential for better generalization in specific out-of-distribution domains.
• Further research is needed to improve deep learning generalization in HAR.
• Domain adaptation techniques may be crucial for robust HAR systems.