Interaction-Aware Transformer Network for Human-Object Interaction Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human-object interaction (HOI) detection aims to identify and classify interactions between humans and objects.
• Current HOI methods often rely on transformer networks and explicit object-level priors.
• Existing approaches overlook the crucial implicit action-level information inherent in HOIs.

Purpose of the Study:

• To propose an interaction-aware transformer network (IATN) for enhanced HOI detection.
• To effectively integrate both implicit action-level and explicit object-level priors.
• To improve the accuracy of joint localization and classification of HOIs.

Main Methods:

• Developed an interaction-aware transformer network (IATN).
• Introduced an action-aware module (AAM) to aggregate scene-level and instance-level action priors.
• Designed an action-oriented graph (AOG) to jointly aggregate action and object level priors.
• Utilized knowledge distillation to enhance action-level priors.

Main Results:

• The proposed IATN effectively utilizes implicit action-level and explicit object-level priors.
• Experiments on HICO-DET and V-COCO datasets demonstrate the model's effectiveness.
• The interaction-aware query acquisition leads to improved HOI predictions.

Conclusions:

• The proposed IATN significantly advances HOI detection by incorporating action-level information.
• The novel approach of jointly utilizing implicit and explicit priors offers a more comprehensive understanding of HOIs.
• IATN provides a robust framework for accurate human-object interaction recognition.