PPDM++: Parallel Point Detection and Matching for Fast and Accurate HOI Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human-Object Interaction (HOI) detection is crucial for understanding human activities.
• Existing two-stage, instance-driven HOI methods suffer from inefficiency due to numerous non-interactive human-object pairs.
• This inefficiency hinders both performance and speed of HOI detection.

Purpose of the Study:

• To propose a novel interaction-driven paradigm for HOI detection.
• To develop a framework that prioritizes interaction detection over instance detection.
• To enhance the efficiency and performance of HOI detection systems.

Main Methods:

• Introduced a top-down, interaction-driven paradigm, contrasting with previous bottom-up, instance-driven approaches.
• Formulated HOI detection as a point triplet: human point, interaction point, object point.
• Developed the Parallel Point Detection and Matching (PPDM) framework and its extension PPDM++ for HOI detection.

Main Results:

• PPDM framework successfully converts the paradigm from instance-driven to interaction-driven, reducing computational overhead.
• PPDM-DLA-34 achieved 19.94 mAP at 42 FPS, establishing the first real-time HOI detector.
• PPDM++-SwinB reached 30.1 mAP at 17 FPS on the HICO-DET dataset.
• Introduced HOI-A, a new application-oriented database for HOI research.

Conclusions:

• The proposed interaction-driven paradigm significantly improves HOI detection efficiency and performance.
• PPDM and PPDM++ offer effective solutions for real-time and high-accuracy HOI detection.
• The development of HOI-A provides a valuable resource for advancing HOI research.