DGIG-Net: Dynamic Graph-in-Graph Networks for Few-Shot Human-Object Interaction.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot learning (FSL) for human-object interaction (HOI) faces challenges due to action diversity and interactivity, leading to unsatisfactory performance with traditional methods.
• Existing FSL approaches struggle to adapt classifiers for ambiguous interclass information in complex HOI scenes.

Purpose of the Study:

• To propose a novel graph prototypes framework, dynamic graph-in-graph networks (DGIG-Net), for few-shot HOI.
• To learn a dynamic metric space by embedding visual subgraphs into task-oriented cross-modal graphs for improved HOI recognition.

Main Methods:

• Constructing a knowledge reconstruction graph to learn latent HOI category representations from visual features.
• Developing a dynamic relation graph integrating visual nodes and semantic information to create a graph metric space for HOI class prototypes.
• Utilizing similarities among actions and objects to extract discriminative information within the graph metric space.

Main Results:

• DGIG-Net demonstrates significant improvements over existing FSL approaches on multiple benchmark datasets.
• The proposed framework achieves state-of-the-art results in few-shot HOI recognition.
• DGIG-Net effectively learns a dynamic metric space, enhancing the ability to handle complex HOI scenes.

Conclusions:

• DGIG-Net offers a novel and effective solution for few-shot human-object interaction recognition.
• The graph prototypes framework successfully addresses the limitations of traditional FSL methods in complex HOI scenarios.
• The proposed approach sets a new standard for performance in few-shot HOI tasks.