Attention-Based Variational Autoencoder Models for Human-Human Interaction Recognition via Generation.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Human intent prediction is vital for development and has numerous applications.
• Existing models for intent prediction are limited.
• Early development of intent prediction highlights its importance.

Purpose of the Study:

• Propose novel attention-based agent models for predicting intent from interacting 3D skeletons.
• Develop inherently multimodal models with perceptual and proprioceptive pathways.
• Improve the efficiency and accuracy of intent prediction models.

Main Methods:

• Utilized two attention-based agent models that sample 3D skeletons via glimpses.
• Incorporated multimodal pathways (perceptual and proprioceptive) into agent designs.
• Trained agents to minimize generation and classification errors by learning optimal sampling strategies.

Main Results:

• One proposed model achieved classification and generation accuracies comparable to state-of-the-art.
• The effective model contained fewer trainable parameters than existing methods.
• Evaluations were performed on benchmark datasets against a state-of-the-art model.

Conclusions:

• The proposed multimodal agent models show promise for accurate and efficient intent prediction.
• Insights from these models can guide the development of future artificial intelligence agents.
• The findings contribute to advancements in human-robot interaction and autonomous systems.