Semantic-Aware Dynamic Generation Networks for Few-Shot Human-Object Interaction Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human-object interaction (HOI) recognition is crucial for understanding complex visual scenes.
• Existing HOI methods face challenges with the long-tail distribution of interactions and the combinatorial explosion of action-object pairs.

Purpose of the Study:

• To address the limitations of current HOI recognition models.
• To propose a novel few-shot learning approach for HOI recognition.
• To develop a dynamic generation method that enhances HOI recognition performance.

Main Methods:

• Introduced Semantic-Aware Dynamic Generation Networks (SADG-Nets) for few-shot HOI recognition.
• Developed a method to assign semantic-aware task representations and generate dynamic parameters.
• Designed a dual semantic-aware encoder module (DSAE-Module) with verb-aware and noun-aware branches.

Main Results:

• SADG-Nets achieved superior performance on few-shot HOI recognition benchmarks.
• The proposed method demonstrated effectiveness in handling the long-tail and combinatorial explosion problems.
• Experimental results on HICO-FS and TUHOI-FS datasets validate the approach's superiority.

Conclusions:

• SADG-Nets offer an effective solution for few-shot human-object interaction recognition.
• The dynamic generation and dual semantic-aware encoding contribute to improved intercategory discriminability and intracategory commonality.
• The approach generalizes well to novel interaction combinations.